{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-08-11T09:21:12.743Z","headline":"LLM 0.32 发布：新增推理轨迹、OpenAI Responses、服务端工具与更智能的日志","description":"Simon Willison 发布 LLM 0.32，这是该项目自启动以来最重要的新版本。新版本支持显示推理轨迹、服务端工具、OpenAI Responses API，并默认使用 GPT-5.6 Luna 模型。","url":"https://www.aioga.com/news/cmsfdkkcs1ocxro2ed4z3s131/","mainEntityOfPage":"https://www.aioga.com/news/cmsfdkkcs1ocxro2ed4z3s131/","datePublished":"2026-08-04T23:58:24.000Z","dateModified":"2026-08-04T23:58:24.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://simonwillison.net/2026/Aug/4/new-release-of-llm","https://aihot.virxact.com/items/cmsfdkkcs1ocxro2ed4z3s131"],"canonicalUrl":"https://www.aioga.com/news/cmsfdkkcs1ocxro2ed4z3s131/","directAnswer":{"@type":"Answer","text":"LLM 0.32 发布，作者称其为项目启动以来最重要的一次版本更新。版本加入可见推理轨迹、服务端提供商工具、重新设计的内容寻址 SQLite 日志，并支持由 OpenAI Responses API 带来的新能力。","url":"https://www.aioga.com/news/cmsfdkkcs1ocxro2ed4z3s131/","dateCreated":"2026-08-04T23:58:24.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"Simon Willison 博客 source article","url":"https://simonwillison.net/2026/Aug/4/new-release-of-llm","datePublished":"2026-08-04T23:58:24.000Z","provider":{"@type":"Organization","name":"Simon Willison 博客","url":"https://simonwillison.net/2026/Aug/4/new-release-of-llm"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.virxact.com/items/cmsfdkkcs1ocxro2ed4z3s131","datePublished":"2026-08-04T23:58:24.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.virxact.com/items/cmsfdkkcs1ocxro2ed4z3s131"}}],"aggregationSource":"Simon Willison 博客","originalPublisher":{"name":"Simon Willison 博客","url":"https://simonwillison.net/2026/Aug/4/new-release-of-llm"},"geoDeepAnswer":null,"article":{"id":"cmsfdkkcs1ocxro2ed4z3s131","slug":"cmsfdkkcs1ocxro2ed4z3s131","url":"https://www.aioga.com/news/cmsfdkkcs1ocxro2ed4z3s131/","title":"LLM 0.32 发布：新增推理轨迹、OpenAI Responses、服务端工具与更智能的日志","title_en":"New release of LLM adds support for reasoning traces， OpenAI Responses， server-side tools， and smarter logging","summary":"Simon Willison 发布 LLM 0.32，这是该项目自启动以来最重要的新版本。新版本支持显示推理轨迹、服务端工具、OpenAI Responses API，并默认使用 GPT-5.6 Luna 模型。","source":"Simon Willison 博客","sourceUrl":"https://simonwillison.net/2026/Aug/4/new-release-of-llm","aiHotUrl":"https://aihot.virxact.com/items/cmsfdkkcs1ocxro2ed4z3s131","publishedAt":"2026-08-04T23:58:24.000Z","category":"产品更新","score":79,"selected":true,"articleBody":["I released LLM 0.32：https://llm.datasette.io/en/stable/changelog.html#v0-32 this morning, the most significant new version of LLM since the initial launch of the project. The new version includes support for visible reasoning traces, server-side provider tools, redesigned content-addressable SQLite logs, new models, and new features enabled by the OpenAI Responses API. I also released a new version of the llm-anthropic plugin：https://github.com/simonw/llm-anthropic with substantial updates of its own.","Running LLM against reasoning models now displays their reasoning traces to standard error, so you can see what they are “thinking” without that information being included in the standard output that you might pipe to another tool. Add -R/--hide-reasoning to turn this off.","LLM includes support out-of-the-box for the GPT-5.6 model family , and the new default model used with llm \"prompt\" is now the inexpensive but capable GPT-5.6 Luna .","LLM calls can now use server-side tools from various providers. OpenAI provide a code execution environment：https://llm.datasette.io/en/stable/openai-models.html#code-interpreter as a server-side tool; LLM can now run prompts that benefit from that like so:","OpenAI also gets a WebSearch：https://llm.datasette.io/en/stable/openai-models.html#web-search tool.","The llm-anthropic：https://github.com/simonw/llm-anthropic plugin adds WebSearch：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search, WebFetch：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch, CodeExecution：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution, and AnthropicMCP：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector, which looks like this:","That causes Anthropic to execute MCP calls against my new datasette-mcp：https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp plugin as part of a single request/response interaction with their API.","The new llm openai endpoint command provides a tool for executing prompts against any OpenAI compatible endpoint：https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it as a one-liner. These aren’t logged, which makes this a handy tool for running one-off prompts against anything that speaks the lingua franca of the LLM API world.","Here’s how I use that to run prompts against Gemma 4 12B running in my localhost LM Studio：https://lmstudio.ai API, via uvx (no LLM installation required) and mixing in the llm-tools-quickjs：https://github.com/simonw/llm-tools-quickjs tool plugin for good measure:","LLM’s Python API previously required you to create a conversation and then send messages to it one at a time. This was an abstraction over the true nature of LLMs, where each request carries a complete history of the messages that came before it. That abstraction started to get in the way for some more advanced cases, so the new release introduces a model.prompt(messages=[]) parameter that can be used like this:","LLM previously returned an iterable sequence of strings from each prompt. This worked great when models returned a string response, but failed to predict the weird shape that models would evolve towards. Today many models return a mix of reasoning text, output strings, tool calls, and even image attachments. With LLM 0.32 you can do this instead：https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events:","Combine these features and we can finally provide a robust implementation of the semi-standard OpenAI chat completions API, which I’ve now released as the llm-chat-completions-server：https://github.com/simonw/llm-chat-completions-server plugin:","Now you can run prompts against LLM via that server, using the new llm openai endpoint command!","The bigger challenge with that kind of API concerns logging. If we’re going to support the pattern where the message sequence is appended to on every request, ideally we can avoid logging all of that duplicate JSON for every turn.","The solution is the new content-addressable message store：https://llm.datasette.io/en/stable/logging.html#the-message-store, modeled after Git. You can see the new schema for that in the documentation：https://llm.datasette.io/en/stable/logging.html#sql-schema, but the llm logs and llm logs --json commands have both been upgraded to convert that format back into something that’s easy to consume.","There is a whole lot more in this release. The 0.32 release notes：https://llm.datasette.io/en/stable/changelog.html#v0-32 are pretty comprehensive, and the notes for 0.32rc2：https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30, 0.32rc：https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30, 0.32a3：https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09, 0.32a2：https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12, and 0.32a0：https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 should fill in any gaps.","Existing LLM plugins should all continue to work, but plugins that provide extra models will need to be upgraded to 0.32 in order to participate fully in the new streaming events system. There’s a guide to implementing plugins with Structured messages and streaming events：https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events in the documentation.","I’ve updated some of my own plugins:","Quite a few of the lower-level tools changes in this release were driven by the needs of Datasette Agent：https://agent.datasette.io/. When I started work on LLM, the term “agent” had such a vague definition that I refused to use it. In September 2025：https://simonwillison.net/2025/Sep/18/agents/ I came around to the idea that \" An LLM agent runs tools in a loop to achieve a goal \" is well established enough now that I could stop avoiding the term entirely.","Tool chains can now pause for human approval：https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause and resume from a stored message history：https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume—both needed by Datasette Agent.","Looking at LLM today it’s beginning to look very agent-shaped to me. There’s something neat about having a CLI utility that can mix and match different tools from different sources with different models all as a one-liner, and that includes a Python library powerful enough to build systems like Datasette Agent：https://agent.datasette.io/ and llm-coding-agent：https://github.com/simonw/llm-coding-agent.","Maybe the next version of LLM will bake the concept of an “agent” into the core library. I’m still trying to figure out what that would look like.","This is New release of LLM adds support for reasoning traces, OpenAI Responses, server-side tools, and smarter logging by Simon Willison, posted on 4th August 2026：/2026/Aug/4/.","Part of series New releases of LLM：/series/llm-releases/","Previous: Stateless MCP has recaptured my interest (and inspired mcp-explorer and datasette-mcp)：/2026/Jul/31/stateless-mcp/","Sponsor me for $10/month and get a curated email digest of the month's most important LLM developments."],"articleImages":[{"sourceUrl":"https://static.simonwillison.net/static/2026/best-pelicans.gif","alt":"Running llm \"think about the best thing about pelicans\" in the macOS terminal window - grey text outputs saying Exploring pelican qualities, then after a paragraph of that a white paragraph of text comes out saying: The best thing about pelicans is their wonderfully oversized, practical design: that enormous bill and pouch look comical, but they make pelicans remarkably skilled fishers. Even better, many species cooperate—working together to herd fish before scooping them up. They’re a great mix of goofy, graceful, and surprisingly clever.","afterParagraph":1,"url":"/media/articles/cmsfdkkcs1ocxro2ed4z3s131/1b61c9606f9e4793.gif"},{"sourceUrl":"https://static.simonwillison.net/static/2026/openai-endpoint-gemma.webp","alt":"Output reads Tool call: QuickJS_execute_javascript({'javascript': '3434 * 2434'}) 8358356 The result of 3434 * 2434 is 8,358,356.","afterParagraph":8,"url":"/media/articles/cmsfdkkcs1ocxro2ed4z3s131/3064b0456beade70.webp"}],"mediaStatus":"ok","articleBodyZh":["我今天早上发布了 LLM 0.32：https://llm.datasette.io/en/stable/changelog.html#v0-32，这是 LLM 自项目初次发布以来最重要的新版本。新版本包括对可见推理轨迹的支持、服务器端提供者工具、重新设计的内容寻址 SQLite 日志、新模型以及通过 OpenAI Responses API 启用的新功能。我还发布了 llm-anthropic 插件的新版本：https://github.com/simonw/llm-anthropic，自身也有大量更新。","在推理模型上运行 LLM 现在会将它们的推理轨迹显示到标准错误输出，因此你可以看到它们的“思考”内容，而这些信息不会包含在你可能传给其他工具的标准输出中。使用 -R/--hide-reasoning 可以关闭此功能。","LLM 开箱即用地支持 GPT-5.6 模型系列，并且使用 llm \"prompt\" 的新默认模型现在是廉价但功能强大的 GPT-5.6 Luna。","LLM 调用现在可以使用来自各种提供者的服务器端工具。OpenAI 提供了一个代码执行环境：https://llm.datasette.io/en/stable/openai-models.html#code-interpreter 作为服务器端工具；LLM 现在可以运行利用该工具受益的提示，如下所示：","OpenAI 还获得了一个 WebSearch：https://llm.datasette.io/en/stable/openai-models.html#web-search 工具。","llm-anthropic：https://github.com/simonw/llm-anthropic 插件增加了 WebSearch：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search、WebFetch：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch、CodeExecution：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution，以及 AnthropicMCP：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector，看起来是这样的：","这样会导致 Anthropic 在与其 API 的单次请求/响应交互中，对我的新 datasette-mcp：https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp 插件执行 MCP 调用。","新的 llm openai endpoint 命令提供了一个工具，可以对任何兼容 OpenAI 的端点执行提示：https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it，一行命令即可完成。这些操作不会被记录，因此它是一个便捷的工具，可以对任何使用 LLM API 世界通用语言的对象运行一次性提示。","这是我如何使用它，通过 uvx（无需安装 LLM）并结合 llm-tools-quickjs：https://github.com/simonw/llm-tools-quickjs 工具插件，对运行在我本地主机 LM Studio：https://lmstudio.ai 的 Gemma 4 12B 进行提示的方式：","LLM 的 Python API 以前要求你创建一个会话，然后一次发送一条消息。这是对 LLM 真正特性的抽象，每个请求都携带之前所有消息的完整历史。对于一些更高级的情况，这种抽象开始成为障碍，因此新版本引入了 model.prompt(messages=[]) 参数，可以这样使用：","LLM 以前会从每个提示返回一个字符串的可迭代序列。当模型返回字符串响应时，这效果很好，但无法预测模型会演变出的奇怪形式。如今，许多模型返回推理文本、输出字符串、工具调用，甚至图像附件的混合内容。在 LLM 0.32 中，你可以改用这个方法：https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events","结合这些功能，我们终于可以提供一个稳健的半标准 OpenAI 聊天完成 API 实现，我已经将其作为 llm-chat-completions-server：https://github.com/simonw/llm-chat-completions-server 插件发布了：","现在，你可以通过该服务器使用新的 llm openai endpoint 命令来对 LLM 运行提示！","这种类型的 API 更大的挑战在于日志记录。如果我们要支持每次请求都追加消息序列的模式，理想情况下可以避免为每次会话记录所有重复的 JSON。","解决方案是新的内容可寻址消息存储：https://llm.datasette.io/en/stable/logging.html#the-message-store， 模仿 Git 建模。你可以在文档中看到新的模式：https://llm.datasette.io/en/stable/logging.html#sql-schema，但 llm logs 和 llm logs --json 命令都已升级，可以将该格式转换回易于使用的形式。","本次发布还有更多内容。0.32 发布说明：https://llm.datasette.io/en/stable/changelog.html#v0-32 相当全面，0.32rc2：https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30、0.32rc：https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30、0.32a3：https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09、0.32a2：https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12 和 0.32a0：https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 的说明可以填补任何遗漏。","现有的 LLM 插件应该都可以继续使用，但提供额外模型的插件需要升级到 0.32 才能完全参与新的流事件系统。文档中有关于使用结构化消息和流事件实现插件的指南：https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events。","我更新了一些我自己的插件：","本次发布的许多底层工具更改是由 Datasette Agent：https://agent.datasette.io/ 的需求驱动的。当我开始开发 LLM 时，“代理”这个术语定义非常模糊，我当时拒绝使用它。到 2025 年 9 月：https://simonwillison.net/2025/Sep/18/agents/ 我接受了这样一个观点：“LLM 代理通过循环运行工具以实现目标” 已经足够成熟，我可以完全停止回避这个术语。","工具链现在可以暂停以等待人工批准：https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause，并从存储的消息历史中恢复：https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume——这两点都是 Datasette Agent 所需的。","今天看LLM，它开始在我看来呈现出非常像代理（agent）的形态。有一个很妙的地方是，你可以用一个命令行工具将来自不同来源、使用不同模型的各种工具混合搭配，一行代码就能搞定，而且它包括一个足够强大的Python库，可以构建像Datasette Agent（https://agent.datasette.io/）和llm-coding-agent（https://github.com/simonw/llm-coding-agent）这样的系统。","也许下一版本的LLM会将“代理”的概念直接融入核心库。我仍在尝试弄清楚那会是什么样子。","这是LLM的新版本，由Simon Willison在2026年8月4日发布：新增对推理追踪、OpenAI响应、服务器端工具以及更智能日志记录的支持：/2026/Aug/4/。","系列文章“大语言模型（LLM）新版本发布”的一部分：/series/llm-releases/","上一篇：无状态MCP重新引起了我的兴趣（并启发了mcp-explorer和datasette-mcp）：/2026/Jul/31/stateless-mcp/","每月赞助我10美元，即可收到精选邮件摘要，汇总当月最重要的LLM动态。"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"LLM 0.32 发布，作者称其为项目启动以来最重要的一次版本更新。版本加入可见推理轨迹、服务端提供商工具、重新设计的内容寻址 SQLite 日志，并支持由 OpenAI Responses API 带来的新能力。","background":"LLM 是 Simon Willison 发布的命令行工具项目。本次更新同时覆盖核心工具与 llm-anthropic 插件：核心工具新增模型、工具调用和日志改造，插件则加入网页搜索、网页抓取、代码执行及 Anthropic MCP 连接器等支持。","viewpoint":"Aioga 判断，这次更新的重点不只是接入新模型，而是把推理展示、服务端工具调用与本地日志管理整合进命令行工作流。推理轨迹默认写入标准错误而非标准输出的设计，可能有助于兼顾观察过程与管道处理结果。","implications":"LLM 现可开箱使用 GPT-5.6 模型家族，且 llm \"prompt\" 的默认模型改为 GPT-5.6 Luna。OpenAI 侧可使用代码执行和网页搜索工具；Anthropic 插件也扩展了多类服务端工具，值得关注其对多提供商工作流的影响。","nextStep":"用户可评估现有脚本对默认模型变更的适配情况，并测试推理轨迹在标准错误中的输出行为。对于一次性调用，新的 llm openai endpoint 命令可面向兼容 OpenAI API 的端点执行提示词；原文明确说明此类调用不会被记录。","evidenceRefs":["title","summary","articleBody","source"],"status":"published","aiGenerated":true,"autoApproved":true,"generatedBy":"aioga-editorial:gpt-5.6-sol","reviewedBy":"aioga-editorial-review:gpt-5.6-sol","generatedAt":"2026-08-05T01:24:30.521Z","sourceHash":"689674bb80baeb48","review":{"approved":true,"groundedness":97,"clarity":92,"duplicationRisk":18,"blockingIssues":[],"notes":["候选内容中的版本功能、默认模型、推理轨迹输出位置、服务端工具、Anthropic 插件能力及 llm openai endpoint 不记录调用等信息均有来源材料支持。","“可能有助于兼顾观察过程与管道处理结果”和“值得关注其对多提供商工作流的影响”属于明确标示且经过限定的分析性判断，未冒充来源事实。"]},"validation":{"passed":true,"mode":"ai-auto","revisions":0,"checks":["schema","length","source-attribution","low-source-overlap","no-html","independent-ai-review"]}},"tags":["产品更新","Simon Willison 博客"],"translations":{"zh-CN":{"title":"LLM 0.32 发布：新增推理轨迹、OpenAI Responses、服务端工具与更智能的日志","summary":"Simon Willison 发布 LLM 0.32，这是该项目自启动以来最重要的新版本。新版本支持显示推理轨迹、服务端工具、OpenAI Responses API，并默认使用 GPT-5.6 Luna 模型。","category":"产品更新","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"LLM 0.32 发布：新增推理轨迹、OpenAI Responses、服务端工具与更智能的日志 - Aioga AI资讯","description":"Simon Willison 发布 LLM 0.32，这是该项目自启动以来最重要的新版本。新版本支持显示推理轨迹、服务端工具、OpenAI Responses API，并默认使用 GPT-5.6 Luna 模型。","url":"https://www.aioga.com/news/cmsfdkkcs1ocxro2ed4z3s131/","articleBody":["我今天早上发布了 LLM 0.32：https://llm.datasette.io/en/stable/changelog.html#v0-32，这是 LLM 自项目初次发布以来最重要的新版本。新版本包括对可见推理轨迹的支持、服务器端提供者工具、重新设计的内容寻址 SQLite 日志、新模型以及通过 OpenAI Responses API 启用的新功能。我还发布了 llm-anthropic 插件的新版本：https://github.com/simonw/llm-anthropic，自身也有大量更新。","在推理模型上运行 LLM 现在会将它们的推理轨迹显示到标准错误输出，因此你可以看到它们的“思考”内容，而这些信息不会包含在你可能传给其他工具的标准输出中。使用 -R/--hide-reasoning 可以关闭此功能。","LLM 开箱即用地支持 GPT-5.6 模型系列，并且使用 llm \"prompt\" 的新默认模型现在是廉价但功能强大的 GPT-5.6 Luna。","LLM 调用现在可以使用来自各种提供者的服务器端工具。OpenAI 提供了一个代码执行环境：https://llm.datasette.io/en/stable/openai-models.html#code-interpreter 作为服务器端工具；LLM 现在可以运行利用该工具受益的提示，如下所示：","OpenAI 还获得了一个 WebSearch：https://llm.datasette.io/en/stable/openai-models.html#web-search 工具。","llm-anthropic：https://github.com/simonw/llm-anthropic 插件增加了 WebSearch：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search、WebFetch：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch、CodeExecution：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution，以及 AnthropicMCP：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector，看起来是这样的：","这样会导致 Anthropic 在与其 API 的单次请求/响应交互中，对我的新 datasette-mcp：https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp 插件执行 MCP 调用。","新的 llm openai endpoint 命令提供了一个工具，可以对任何兼容 OpenAI 的端点执行提示：https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it，一行命令即可完成。这些操作不会被记录，因此它是一个便捷的工具，可以对任何使用 LLM API 世界通用语言的对象运行一次性提示。","这是我如何使用它，通过 uvx（无需安装 LLM）并结合 llm-tools-quickjs：https://github.com/simonw/llm-tools-quickjs 工具插件，对运行在我本地主机 LM Studio：https://lmstudio.ai 的 Gemma 4 12B 进行提示的方式：","LLM 的 Python API 以前要求你创建一个会话，然后一次发送一条消息。这是对 LLM 真正特性的抽象，每个请求都携带之前所有消息的完整历史。对于一些更高级的情况，这种抽象开始成为障碍，因此新版本引入了 model.prompt(messages=[]) 参数，可以这样使用：","LLM 以前会从每个提示返回一个字符串的可迭代序列。当模型返回字符串响应时，这效果很好，但无法预测模型会演变出的奇怪形式。如今，许多模型返回推理文本、输出字符串、工具调用，甚至图像附件的混合内容。在 LLM 0.32 中，你可以改用这个方法：https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events","结合这些功能，我们终于可以提供一个稳健的半标准 OpenAI 聊天完成 API 实现，我已经将其作为 llm-chat-completions-server：https://github.com/simonw/llm-chat-completions-server 插件发布了：","现在，你可以通过该服务器使用新的 llm openai endpoint 命令来对 LLM 运行提示！","这种类型的 API 更大的挑战在于日志记录。如果我们要支持每次请求都追加消息序列的模式，理想情况下可以避免为每次会话记录所有重复的 JSON。","解决方案是新的内容可寻址消息存储：https://llm.datasette.io/en/stable/logging.html#the-message-store， 模仿 Git 建模。你可以在文档中看到新的模式：https://llm.datasette.io/en/stable/logging.html#sql-schema，但 llm logs 和 llm logs --json 命令都已升级，可以将该格式转换回易于使用的形式。","本次发布还有更多内容。0.32 发布说明：https://llm.datasette.io/en/stable/changelog.html#v0-32 相当全面，0.32rc2：https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30、0.32rc：https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30、0.32a3：https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09、0.32a2：https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12 和 0.32a0：https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 的说明可以填补任何遗漏。","现有的 LLM 插件应该都可以继续使用，但提供额外模型的插件需要升级到 0.32 才能完全参与新的流事件系统。文档中有关于使用结构化消息和流事件实现插件的指南：https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events。","我更新了一些我自己的插件：","本次发布的许多底层工具更改是由 Datasette Agent：https://agent.datasette.io/ 的需求驱动的。当我开始开发 LLM 时，“代理”这个术语定义非常模糊，我当时拒绝使用它。到 2025 年 9 月：https://simonwillison.net/2025/Sep/18/agents/ 我接受了这样一个观点：“LLM 代理通过循环运行工具以实现目标” 已经足够成熟，我可以完全停止回避这个术语。","工具链现在可以暂停以等待人工批准：https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause，并从存储的消息历史中恢复：https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume——这两点都是 Datasette Agent 所需的。","今天看LLM，它开始在我看来呈现出非常像代理（agent）的形态。有一个很妙的地方是，你可以用一个命令行工具将来自不同来源、使用不同模型的各种工具混合搭配，一行代码就能搞定，而且它包括一个足够强大的Python库，可以构建像Datasette Agent（https://agent.datasette.io/）和llm-coding-agent（https://github.com/simonw/llm-coding-agent）这样的系统。","也许下一版本的LLM会将“代理”的概念直接融入核心库。我仍在尝试弄清楚那会是什么样子。","这是LLM的新版本，由Simon Willison在2026年8月4日发布：新增对推理追踪、OpenAI响应、服务器端工具以及更智能日志记录的支持：/2026/Aug/4/。","系列文章“大语言模型（LLM）新版本发布”的一部分：/series/llm-releases/","上一篇：无状态MCP重新引起了我的兴趣（并启发了mcp-explorer和datasette-mcp）：/2026/Jul/31/stateless-mcp/","每月赞助我10美元，即可收到精选邮件摘要，汇总当月最重要的LLM动态。"]},"en":{"title":"LLM 0.32 released: added inference trajectories, OpenAI Responses, server-side tools, and smarter logging","summary":"Simon Willison released LLM 0.32, the most important new version since the project's launch. The new version supports displaying inference trajectories, server-side tools, the OpenAI Responses API, and uses the GPT-5.6 Luna model by default.","category":"Products","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"LLM 0.32 released: added inference trajectories, OpenAI Responses, server-side tools, and smarter logging - Aioga AI News","description":"Simon Willison released LLM 0.32, the most important new version since the project's launch. The new version supports displaying inference trajectories, server-side tools, the Open...","url":"https://www.aioga.com/en/news/cmsfdkkcs1ocxro2ed4z3s131/","contentTranslated":true,"sourceHash":"a2973b505f038783","translatedAt":"2026-08-05T01:22:34.836Z","articleBody":["This morning I released LLM 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32, which is the most important new version of LLM since the project was first launched. The new version includes support for visible reasoning traces, server-side provider tools, a redesigned content-addressable SQLite log, new models, and new features enabled through the OpenAI Responses API. I also released a new version of the llm-anthropic plugin: https://github.com/simonw/llm-anthropic, which itself has many updates.","Running LLMs on reasoning models now outputs their reasoning traces to standard error, so you can see their 'thoughts,' which are not included in the standard output that you might pass to other tools. You can use -R/--hide-reasoning to disable this feature.","LLMs natively support the GPT-5.6 model series out of the box, and the new default model for using llm 'prompts' is now the inexpensive but powerful GPT-5.6 Luna.","LLM calls can now use server-side tools from various providers. OpenAI provides a code execution environment: https://llm.datasette.io/en/stable/openai-models.html#code-interpreter as a server-side tool; LLMs can now run prompts that benefit from using this tool, as shown below:","OpenAI also obtained a WebSearch tool: https://llm.datasette.io/en/stable/openai-models.html#web-search.","llm-anthropic: https://github.com/simonw/llm-anthropic The plugin adds WebSearch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search, WebFetch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch, CodeExecution: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution, and AnthropicMCP: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector, and it looks like this:","This will cause Anthropic to make MCP calls on my new datasette-mcp: https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp plugin during single request/response interactions with its API.","The new llm OpenAI endpoint command provides a tool that can execute prompts on any OpenAI-compatible endpoint: https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it, all in a single command. These operations are not logged, making it a convenient tool for running one-time prompts on any object that uses the LLM API world's universal language.","Here's how I use it: via uvx (no need to install LLM) and combined with llm-tools-quickjs: https://github.com/simonw/llm-tools-quickjs tool plugin, I use Gemma 4 12B running on my local host LM Studio: https://lmstudio.ai:","The Python API for LLM used to require you to create a session and then send one message at a time. This is an abstraction of the true characteristics of LLM, where each request carries the full history of all previous messages. For some more advanced scenarios, this abstraction begins to become an obstacle, so the new version introduces the model.prompt(messages=[]) parameter, which can be used like this:","LLMs used to return an iterable sequence of strings from each prompt. This worked well when the model returned string responses, but it couldn't predict the strange forms the model might evolve. Nowadays, many models return a mix of reasoning text, output strings, tool calls, and even image attachments. In LLM 0.32, you can switch to using this method: https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events","Combining these features, we can finally provide a robust semi-standard OpenAI chat completions API implementation, which I have released as the llm-chat-completions-server plugin: https://github.com/simonw/llm-chat-completions-server","Now, you can use the new llm openai endpoint command through this server to run prompts on the LLM!","The bigger challenge for this type of API lies in logging. If we want to support a mode where each request appends to the message sequence, ideally we can avoid logging all the repeated JSON for each session.","The solution is a new content-addressable message store: https://llm.datasette.io/en/stable/logging.html#the-message-store, modeled after Git. You can see the new schema in the documentation: https://llm.datasette.io/en/stable/logging.html#sql-schema, and both the llm logs and llm logs --json commands have been upgraded to convert this format back into an easy-to-use form.","There is more content in this release. The 0.32 release notes: https://llm.datasette.io/en/stable/changelog.html#v0-32 are quite comprehensive; 0.32rc2: https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30, 0.32rc: https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30, 0.32a3: https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09, 0.32a2: https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12, and 0.32a0: https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 notes can fill in any omissions.","Existing LLM plugins should all continue to work, but plugins that provide additional models need to be upgraded to 0.32 to fully participate in the new stream event system. The documentation has a guide on implementing plugins using structured messages and stream events: https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events.","I updated some of my own plugins:","Many of the underlying tool changes in this release were driven by the needs of Datasette Agent: https://agent.datasette.io/. When I started developing LLMs, the definition of the term 'agent' was very vague, and I refused to use it at the time. By September 2025: https://simonwillison.net/2025/Sep/18/agents/ I had accepted the view that 'LLM agents achieve goals by running tools in loops', which was mature enough for me to completely stop avoiding the term.","The toolchain can now pause to wait for manual approval: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause, and resume from stored message history: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume — both of these are required by Datasette Agent.","Looking at LLM today, it has begun to appear very much like an agent in my view. One very clever aspect is that you can use a command-line tool to mix and match various tools from different sources using different models, and it can be done with a single line of code. Moreover, it includes a sufficiently powerful Python library that allows you to build systems like Datasette Agent (https://agent.datasette.io/) and llm-coding-agent (https://github.com/simonw/llm-coding-agent).","Maybe the next version of LLM will directly integrate the concept of 'agents' into the core library. I am still trying to figure out what that would look like.","This is the new version of LLM, released by Simon Willison on August 4, 2026: it adds support for reasoning tracking, OpenAI responses, server-side tools, and smarter logging: /2026/Aug/4/.","Part of the series of articles 'New Releases of Large Language Models (LLM)': /series/llm-releases/","Previous article: Stateless MCP rekindled my interest (and inspired mcp-explorer and datasette-mcp): /2026/Jul/31/stateless-mcp/","Sponsor me $10 per month and you will receive a curated email digest summarizing the most important LLM developments of the month."],"bodyTranslated":true,"bodySourceHash":"180cd773be407152","bodyTranslatedAt":"2026-08-08T12:21:49.278Z"},"ja":{"title":"LLM 0.32リリース:推論軌跡の追加、OpenAI応答、サーバーサイドツール、よりスマートなログ記録","summary":"サイモン・ウィリソンは、プロジェクト開始以来最も重要な新バージョンであるLLM 0.32をリリースしました。 新バージョンは推論軌跡の表示、サーバーサイドツール、OpenAIレスポンスAPIをサポートし、デフォルトでGPT-5.6 Lunaモデルを使用しています。","category":"製品更新","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"LLM 0.32リリース:推論軌跡の追加、OpenAI応答、サーバーサイドツール、よりスマートなログ記録 - Aioga AIニュース","description":"サイモン・ウィリソンは、プロジェクト開始以来最も重要な新バージョンであるLLM 0.32をリリースしました。 新バージョンは推論軌跡の表示、サーバーサイドツール、OpenAIレスポンスAPIをサポートし、デフォルトでGPT-5.6 Lunaモデルを使用しています。","url":"https://www.aioga.com/ja/news/cmsfdkkcs1ocxro2ed4z3s131/","contentTranslated":true,"sourceHash":"a2973b505f038783","translatedAt":"2026-08-05T01:22:36.311Z","articleBody":["私は今朝 LLM 0.32 をリリースしました：https://llm.datasette.io/en/stable/changelog.html#v0-32、これは LLM がプロジェクト初公開以来で最も重要な新バージョンです。新バージョンには可視推論トレースのサポート、サーバーサイドプロバイダーツール、再設計されたコンテンツアドレッシング SQLite ログ、新しいモデル、および OpenAI Responses API を通じて有効化された新機能が含まれています。また、llm-anthropic プラグインの新バージョンもリリースしました：https://github.com/simonw/llm-anthropic、こちらも大幅な更新があります。","推論モデル上で LLM を実行すると、現在それらの推論の軌跡が標準エラー出力に表示されるため、それらの「思考」内容を見ることができますが、これらの情報は他のツールに渡す可能性のある標準出力には含まれません。-R/--hide-reasoning を使用すると、この機能を無効にできます。","LLMは箱から出してすぐにGPT-5.6モデルシリーズをサポートしており、llmの「prompt」で使用する新しいデフォルトモデルは、今では低価格ながら強力なGPT-5.6 Lunaです。","LLM の呼び出しは現在、さまざまなプロバイダーからのサーバーサイドツールを使用できます。OpenAI はサーバーサイドツールとして次のコード実行環境を提供しています：https://llm.datasette.io/en/stable/openai-models.html#code-interpreter；LLM は現在、このツールを活用したプロンプトを実行して利益を得ることができます。","OpenAI は WebSearch：https://llm.datasette.io/en/stable/openai-models.html#web-search ツールも獲得しました。","llm-anthropic：https://github.com/simonw/llm-anthropic プラグインには WebSearch：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search、WebFetch：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch、CodeExecution：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution、および AnthropicMCP：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector が追加されており、見た目は次の通りです：","こうすると、Anthropic はその API との単一のリクエスト/レスポンスのやり取りで、私の新しい datasette-mcp：https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp プラグインに対して MCP 呼び出しを実行することになります。","新しい llm OpenAI エンドポイントのコマンドは、OpenAI に対応する任意のエンドポイントでプロンプトを実行できるツールを提供します：https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it、一行のコマンドで完了します。これらの操作は記録されないため、LLM API を使用する世界共通言語の任意のオブジェクトに対してワンタイムのプロンプトを実行する便利なツールです。","これは私がそれを使う方法で、uvx（LLMのインストール不要）を使い、llm-tools-quickjs：https://github.com/simonw/llm-tools-quickjs ツールプラグインと組み合わせて、私のローカルホストで動作する LM Studio：https://lmstudio.ai の Gemma 4 12B にプロンプトを与える方法です：","LLMのPython APIは以前、セッションを作成してから一度に1つのメッセージを送信することを要求していました。これはLLMの本来の特性を抽象化したもので、各リクエストにはこれまでのすべてのメッセージ履歴が含まれています。いくつかのより高度なケースでは、この抽象化が障害となり始めるため、新しいバージョンでは model.prompt(messages=[]) パラメータが導入され、次のように使用できます：","以前の LLM は、各プロンプトから文字列のイテラブルなシーケンスを返していました。モデルが文字列の応答を返す場合、これはうまく機能しましたが、モデルがどのように奇妙な形式に進化するかは予測できませんでした。現在、多くのモデルは、推論テキスト、出力文字列、ツール呼び出し、さらには画像添付を含む混合コンテンツを返します。LLM 0.32 では、この方法に切り替えることができます：https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events","これらの機能を組み合わせることで、ついに堅牢な半標準の OpenAI チャット完了 API 実装を提供できるようになりました。私はこれを llm-chat-completions-server としてプラグインとして公開しました：https://github.com/simonw/llm-chat-completions-server","今、あなたはこのサーバーを通じて新しい LLM OpenAI エンドポイントコマンドを使用して LLM にプロンプトを実行することができます！","このタイプの API のより大きな課題はログ記録にあります。もし私たちが毎回のリクエストでメッセージシーケンスを追加するパターンをサポートする場合、理想的には各セッションのすべての重複した JSON を記録することを避けることができます。","解決策は新しいコンテンツアドレッサブルメッセージストアです：https://llm.datasette.io/en/stable/logging.html#the-message-store、Git のモデリングを模倣しています。ドキュメントでは新しいスキーマを見ることができます：https://llm.datasette.io/en/stable/logging.html#sql-schema、しかし llm logs と llm logs --json コマンドはすでにアップグレードされており、この形式を使いやすい形に変換できます。","今回のリリースにはさらに多くの内容があります。0.32 リリースノート：https://llm.datasette.io/en/stable/changelog.html#v0-32 はかなり包括的で、0.32rc2：https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30、0.32rc：https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30、0.32a3：https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09、0.32a2：https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12、そして 0.32a0：https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 の説明があれば、どんな抜けも補うことができます。","既存の LLM プラグインは引き続き使用できますが、追加のモデルを提供するプラグインは、新しいストリームイベントシステムに完全に対応するために 0.32 にアップグレードする必要があります。文書には、構造化メッセージとストリームイベントを使用してプラグインを実装するためのガイドがあります：https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events。","私は自分のいくつかのプラグインを更新しました：","今回リリースされた多くの低レベルツールの変更は、Datasette Agent：https://agent.datasette.io/ のニーズによって推進されました。私がLLMの開発を始めたとき、「エージェント」という用語の定義は非常に曖昧で、当時は使用を避けていました。2025年9月までには：https://simonwillison.net/2025/Sep/18/agents/ この見解を受け入れるようになりました。「LLMエージェントは目標を達成するためにツールをループで実行する」というものです。私はこの用語の使用を完全に避けるのをやめることができました。","ツールチェーンは現在、手動承認を待つために一時停止することができます：https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause、そして保存されたメッセージ履歴から復元することができます：https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume ——これらの両方は Datasette Agent に必要です。","今日LLMを見て、私にはそれが非常にエージェント（agent）の形態を示し始めているように見えました。非常に面白い点は、異なるソースから、異なるモデルを使ったさまざまなツールをコマンドラインツールで混ぜ合わせることができ、1行のコードでそれを実現できることです。そして、Datasette Agent（https://agent.datasette.io/）やllm-coding-agent（https://github.com/simonw/llm-coding-agent）のようなシステムを構築できる十分に強力なPythonライブラリも含まれています。","おそらく次のバージョンのLLMでは、「エージェント」の概念が直接コアライブラリに統合されるかもしれません。私はまだそれがどのようなものになるのかを理解しようとしています。","これはLLMの新しいバージョンで、Simon Willisonによって2026年8月4日にリリースされました：推論追跡、OpenAIの応答、サーバーサイドツール、そしてより賢いログ記録のサポートが追加されました：/2026/Aug/4/。","シリーズ記事「大規模言語モデル（LLM）新バージョンのリリース」の一部：/series/llm-releases/","前回の記事：ステートレスMCPが再び私の興味を引いた（そしてmcp-explorerとdatasette-mcpにインスピレーションを与えた）：/2026/Jul/31/stateless-mcp/","毎月10ドルを寄付すると、月ごとの最も重要なLLMの動向をまとめた厳選メールサマリーを受け取ることができます。"],"bodyTranslated":true,"bodySourceHash":"180cd773be407152","bodyTranslatedAt":"2026-08-08T12:21:47.252Z"},"ko":{"title":"LLM 0.32 출시: 추론 궤적, OpenAI 응답, 서버 측 도구, 그리고 더 똑똑한 로깅 기능","summary":"사이먼 윌리슨은 프로젝트 시작 이후 가장 중요한 신버전인 LLM 0.32를 출시했습니다. 새 버전은 추론 궤적, 서버 측 도구, OpenAI 응답 API를 지원하며 기본적으로 GPT-5.6 Luna 모델을 사용합니다.","category":"제품 업데이트","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"LLM 0.32 출시: 추론 궤적, OpenAI 응답, 서버 측 도구, 그리고 더 똑똑한 로깅 기능 - Aioga AI 뉴스","description":"사이먼 윌리슨은 프로젝트 시작 이후 가장 중요한 신버전인 LLM 0.32를 출시했습니다. 새 버전은 추론 궤적, 서버 측 도구, OpenAI 응답 API를 지원하며 기본적으로 GPT-5.6 Luna 모델을 사용합니다.","url":"https://www.aioga.com/ko/news/cmsfdkkcs1ocxro2ed4z3s131/","contentTranslated":true,"sourceHash":"a2973b505f038783","translatedAt":"2026-08-05T01:22:44.968Z","articleBody":["저는 오늘 아침 LLM 0.32를 출시했습니다: https://llm.datasette.io/en/stable/changelog.html#v0-32, 이는 LLM이 프로젝트 최초로 출시된 이후 가장 중요한 새 버전입니다. 새 버전에는 가시적 추론 경로 지원, 서버 측 제공자 도구, 재설계된 콘텐츠 주소 기반 SQLite 로그, 새 모델 및 OpenAI Responses API를 통해 활성화된 새로운 기능이 포함되어 있습니다. 또한 llm-anthropic 플러그인의 새 버전도 출시했습니다: https://github.com/simonw/llm-anthropic, 자체적으로도 많은 업데이트가 있습니다.","추론 모델에서 LLM을 실행하면 이제 그들의 추론 경로가 표준 오류 출력에 표시되므로, 그들의 \"생각\" 내용을 볼 수 있으며, 이러한 정보는 다른 도구에 전달할 수 있는 표준 출력에는 포함되지 않습니다. -R/--hide-reasoning을 사용하면 이 기능을 끌 수 있습니다.","LLM은 개봉 즉시 GPT-5.6 모델 시리즈를 지원하며, llm \"prompt\"를 사용할 때의 새로운 기본 모델은 이제 저렴하지만 강력한 GPT-5.6 Luna입니다.","LLM 호출은 이제 다양한 제공자의 서버 측 도구를 사용할 수 있습니다. OpenAI는 코드 실행 환경을 제공합니다: https://llm.datasette.io/en/stable/openai-models.html#code-interpreter 서버 측 도구로서; LLM은 이제 다음과 같이 해당 도구를 활용하는 프롬프트를 실행할 수 있습니다:","OpenAI는 또한 WebSearch：https://llm.datasette.io/en/stable/openai-models.html#web-search 도구를 얻었습니다.","llm-anthropic：https://github.com/simonw/llm-anthropic 플러그인은 WebSearch：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search, WebFetch：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch, CodeExecution：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution, 그리고 AnthropicMCP：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector를 추가했으며, 이렇게 생긴 것 같습니다:","이렇게 하면 Anthropic이 자신의 API와의 단일 요청/응답 상호작용에서 내 새로운 datasette-mcp：https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp 플러그인에 대해 MCP 호출을 수행하게 됩니다.","새로운 llm openai 엔드포인트 명령은 OpenAI와 호환되는 모든 엔드포인트에 프롬프트를 실행할 수 있는 도구를 제공합니다: https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it, 한 줄 명령으로 완료할 수 있습니다. 이러한 작업은 기록되지 않으므로, LLM API 세계 공용 언어를 사용하는 어떤 객체에도 일회성 프롬프트를 실행할 수 있는 편리한 도구입니다.","이것은 제가 그것을 사용하는 방법으로, uvx(LLM 설치 불필요)와 llm-tools-quickjs：https://github.com/simonw/llm-tools-quickjs 도구 플러그인을 결합하여, 제 로컬 호스트 LM Studio：https://lmstudio.ai 에서 실행 중인 Gemma 4 12B에 프롬프트를 보내는 방식입니다:","LLM의 Python API는 이전에 세션을 생성한 다음 한 번에 한 메시지를 보내도록 요구했습니다. 이것은 LLM의 실제 기능을 추상화한 것으로, 각 요청에는 이전 모든 메시지의 전체 기록이 포함됩니다. 일부 더 고급 상황에서는 이러한 추상이 장애물이 되기 시작하여 새 버전에서는 model.prompt(messages=[]) 매개변수가 도입되었고, 이렇게 사용할 수 있습니다:","이전에는 LLM이 각 프롬프트에서 문자열의 반복 가능한 시퀀스를 반환했습니다. 모델이 문자열 응답을 반환할 때는 효과적이었지만, 모델이 생성할 이상한 형태를 예측할 수는 없었습니다. 현재 많은 모델이 추론 텍스트, 출력 문자열, 도구 호출, 심지어 이미지 첨부물까지 혼합된 내용을 반환합니다. LLM 0.32에서는 다음 방법으로 대체할 수 있습니다: https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events","이 기능들을 결합하여, 우리는 마침내 안정적인 준표준 OpenAI 채팅 완성 API 구현을 제공할 수 있게 되었으며, 이를 llm-chat-completions-server：https://github.com/simonw/llm-chat-completions-server 플러그인으로 출시했습니다:","이제, 당신은 이 서버를 통해 새로운 llm openai 엔드포인트 명령어로 LLM에 프롬프트를 실행할 수 있습니다!","이 유형의 API에서 더 큰 도전은 로그 기록에 있습니다. 만약 우리가 매번 요청마다 메시지 시퀀스를 추가하는 방식을 지원하려 한다면, 이상적으로는 매 세션마다 모든 반복된 JSON을 기록하는 것을 피할 수 있습니다.","해결책은 새로운 콘텐츠 주소 지정 메시지 저장소입니다: https://llm.datasette.io/en/stable/logging.html#the-message-store, Git 모델을 모방합니다. 문서에서 새로운 스키마를 확인할 수 있습니다: https://llm.datasette.io/en/stable/logging.html#sql-schema, 하지만 llm logs 및 llm logs --json 명령은 모두 업그레이드되어 이 형식을 다시 사용하기 쉬운 형태로 변환할 수 있습니다.","이번 릴리스에는 더 많은 내용이 있습니다. 0.32 릴리스 노트: https://llm.datasette.io/en/stable/changelog.html#v0-32 상당히 포괄적이며, 0.32rc2: https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30, 0.32rc: https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30, 0.32a3: https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09, 0.32a2: https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12 및 0.32a0: https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28의 설명은 누락된 내용을 보충할 수 있습니다.","기존의 LLM 플러그인은 계속 사용할 수 있지만, 추가 모델을 제공하는 플러그인은 새로운 스트리밍 이벤트 시스템을 완전히 사용하려면 0.32로 업그레이드해야 합니다. 문서에는 구조화된 메시지와 스트리밍 이벤트를 사용하여 플러그인을 구현하는 가이드가 있습니다: https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events.","저는 제 자신의 플러그인 몇 가지를 업데이트했습니다：","이번에 발표된 많은 하위 도구 변경은 Datasette Agent：https://agent.datasette.io/ 의 요구에 의해 추진되었습니다. 제가 LLM 개발을 시작했을 때, '에이전트'라는 용어의 정의는 매우 불분명했고, 그때는 사용을 거부했습니다. 2025년 9월까지：https://simonwillison.net/2025/Sep/18/agents/ 저는 다음과 같은 관점을 수용하게 되었습니다. 'LLM 에이전트는 목표를 달성하기 위해 도구를 반복적으로 실행한다'는 것이 충분히 성숙하여, 저는 이제 이 용어를 완전히 회피하지 않아도 되었습니다.","툴체인은 이제 수동 승인을 기다리기 위해 일시 중지할 수 있습니다: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause, 그리고 저장된 메시지 기록에서 복원할 수 있습니다: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume —— 이 두 가지 모두 Datasette Agent에 필요합니다.","오늘 LLM을 보면서, 그것이 제 눈에는 매우 에이전트(agent) 같은 형태를 나타내기 시작하는 것 같습니다. 정말 멋진 점은, 커맨드 라인 도구를 사용하면 다양한 출처와 다양한 모델을 사용하는 여러 도구를 혼합하여 한 줄의 코드로 해결할 수 있다는 것입니다. 그리고 그것은 Datasette Agent(https://agent.datasette.io/)나 llm-coding-agent(https://github.com/simonw/llm-coding-agent)와 같은 시스템을 구축할 수 있는 충분히 강력한 Python 라이브러리를 포함하고 있습니다.","아마 다음 버전의 LLM은 '에이전트' 개념을 직접 핵심 라이브러리에 통합할 것입니다. 저는 아직 그것이 어떤 모습일지 파악하려고 노력 중입니다.","이것은 LLM의 새 버전으로, Simon Willison이 2026년 8월 4일에 발표했습니다: 추론 추적, OpenAI 응답, 서버 측 도구 및 보다 스마트한 로그 기록 지원이 추가되었습니다: /2026/Aug/4/.","연재 글 '대형 언어 모델(LLM) 새 버전 출시'의 일부: /series/llm-releases/","이전 글: 상태 없는 MCP가 다시 내 관심을 끌었다(그리고 mcp-explorer와 datasette-mcp에 영감을 주었다): /2026/Jul/31/stateless-mcp/","매달 저에게 10달러를 후원하시면, 이번 달 가장 중요한 LLM 소식을 정리한 선정된 이메일 요약을 받으실 수 있습니다."],"bodyTranslated":true,"bodySourceHash":"180cd773be407152","bodyTranslatedAt":"2026-08-08T12:22:34.452Z"},"es":{"title":"Lanzamiento del LLM 0.32: añadido trayectorias de inferencia, respuestas OpenAI, herramientas del lado del servidor y logging más inteligente","summary":"Simon Willison lanzó LLM 0.32, la versión nueva más importante desde el lanzamiento del proyecto. La nueva versión soporta la visualización de trayectorias de inferencia, herramientas del lado del servidor, la API OpenAI Responses y utiliza por defecto el modelo Luna GPT-5.6.","category":"Productos","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Lanzamiento del LLM 0.32: añadido trayectorias de inferencia, respuestas OpenAI, herramientas del lado del servidor y logging más inteligente - Aioga Noticias de IA","description":"Simon Willison lanzó LLM 0.32, la versión nueva más importante desde el lanzamiento del proyecto. La nueva versión soporta la visualización de trayectorias de inferencia, herramien...","url":"https://www.aioga.com/es/news/cmsfdkkcs1ocxro2ed4z3s131/","contentTranslated":true,"sourceHash":"a2973b505f038783","translatedAt":"2026-08-05T01:22:45.082Z","articleBody":["Esta mañana publiqué LLM 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32, esta es la nueva versión más importante de LLM desde el primer lanzamiento del proyecto. La nueva versión incluye soporte para rastreo de razonamiento visible, herramientas de proveedor del lado del servidor, un registro SQLite de direccionamiento de contenido rediseñado, nuevos modelos y nuevas funciones habilitadas a través de la API OpenAI Responses. También publiqué una nueva versión del complemento llm-anthropic: https://github.com/simonw/llm-anthropic, que también tiene muchas actualizaciones.","Ejecutar LLM en modelos de razonamiento ahora mostrará sus trayectorias de razonamiento en la salida de error estándar, por lo que puedes ver su contenido de \"pensamiento\", mientras que esta información no se incluirá en la salida estándar que puedas enviar a otras herramientas. Usar -R/--hide-reasoning puede desactivar esta función.","LLM admite de inmediato la serie de modelos GPT-5.6, y el nuevo modelo predeterminado para usar el \"prompt\" de llm ahora es el GPT-5.6 Luna, barato pero potente.","Ahora las llamadas LLM pueden usar herramientas del lado del servidor de diversos proveedores. OpenAI ofrece un entorno de ejecución de código: https://llm.datasette.io/en/stable/openai-models.html#code-interpreter como una herramienta del lado del servidor; ahora LLM puede ejecutar indicaciones que se beneficien del uso de esa herramienta, como se muestra a continuación:","OpenAI también obtuvo una herramienta WebSearch: https://llm.datasette.io/en/stable/openai-models.html#web-search.","llm-anthropic: https://github.com/simonw/llm-anthropic El complemento ha agregado WebSearch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search, WebFetch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch, CodeExecution: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution, y AnthropicMCP: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector, parece ser así:","Esto hará que Anthropic realice llamadas MCP a mi nuevo complemento datasette-mcp: https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp durante las interacciones de solicitud/respuesta única con su API.","El nuevo comando del endpoint llm de OpenAI proporciona una herramienta que permite ejecutar indicaciones en cualquier endpoint compatible con OpenAI: https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it, todo con un solo comando. Estas operaciones no se registran, por lo que es una herramienta conveniente para ejecutar indicaciones únicas en cualquier objeto que utilice el lenguaje universal del mundo de la API LLM.","Esto es cómo lo utilizo, mediante uvx (sin necesidad de instalar LLM) y combinándolo con el complemento de herramientas llm-tools-quickjs: https://github.com/simonw/llm-tools-quickjs, para dar indicaciones a Gemma 4 12B que se ejecuta en mi host local LM Studio: https://lmstudio.ai","La API de Python de LLM solía requerir que crearas una sesión y luego enviaras un mensaje a la vez. Esto es una abstracción de las verdaderas características de LLM, donde cada solicitud lleva todo el historial completo de mensajes anteriores. Para algunos casos más avanzados, esta abstracción comenzaba a convertirse en un obstáculo, por lo que la nueva versión introdujo el parámetro model.prompt(messages=[]) que se puede usar así:","Antes, LLM devolvía una secuencia iterable de cadenas por cada indicación. Esto funcionaba bien cuando el modelo devolvía respuestas en forma de cadena, pero era impredecible ante las formas extrañas que el modelo podía desarrollar. Hoy en día, muchos modelos devuelven una mezcla de texto de razonamiento, cadenas de salida, llamadas a herramientas e incluso archivos adjuntos de imágenes. En LLM 0.32, puedes usar este método en su lugar: https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events","Combinando estas funciones, finalmente podemos ofrecer una implementación robusta y semi-estándar de la API de finalización de chat de OpenAI. Ya lo he publicado como un complemento llamado llm-chat-completions-server: https://github.com/simonw/llm-chat-completions-server","Ahora, puedes usar el nuevo comando de endpoint llm openai a través de este servidor para ejecutar indicaciones en LLM!","El mayor desafío de este tipo de API está en el registro de logs. Si queremos soportar un modelo en el que se agregue la secuencia de mensajes en cada solicitud, lo ideal sería evitar registrar todo el JSON repetido en cada sesión.","La solución es un nuevo almacenamiento de mensajes direccionable por contenido: https://llm.datasette.io/en/stable/logging.html#the-message-store, modelado imitando Git. Puedes ver el nuevo esquema en la documentación: https://llm.datasette.io/en/stable/logging.html#sql-schema, pero los comandos llm logs y llm logs --json ya se han actualizado para convertir este formato de nuevo en una forma fácil de usar.","Esta versión también incluye más contenido. Notas de la versión 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32 bastante completas, 0.32rc2: https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30, 0.32rc: https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30, 0.32a3: https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09, 0.32a2: https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12 y 0.32a0: https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 pueden llenar cualquier omisión.","Los plugins LLM existentes deberían poder seguir utilizándose, pero los plugins que proporcionan modelos adicionales necesitan actualizarse a la versión 0.32 para poder participar completamente en el nuevo sistema de eventos de flujo. La documentación contiene una guía sobre cómo implementar plugins usando mensajes estructurados y eventos de flujo: https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events.","He actualizado algunos de mis propios complementos:","Muchos de los cambios en las herramientas subyacentes lanzados esta vez fueron impulsados por las necesidades de Datasette Agent: https://agent.datasette.io/. Cuando comencé a desarrollar LLM, la definición del término 'agente' era muy vaga, y en ese momento me negué a usarlo. Para septiembre de 2025: https://simonwillison.net/2025/Sep/18/agents/ acepté la opinión de que 'los agentes LLM funcionan en bucle utilizando herramientas para alcanzar objetivos', lo cual ya estaba lo suficientemente desarrollado como para que pudiera dejar de evitar por completo el uso de este término.","La cadena de herramientas ahora puede pausar para esperar la aprobación manual: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause, y reanudarse desde el historial de mensajes almacenado: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume — ambos son requeridos por Datasette Agent.","Hoy miré LLM, y comenzó a mostrarme una forma muy parecida a un agente. Hay algo muy interesante: puedes usar una herramienta de línea de comandos para combinar diversas herramientas provenientes de diferentes fuentes y usando distintos modelos, todo con una sola línea de código, y además incluye una biblioteca de Python lo suficientemente potente como para construir sistemas como Datasette Agent (https://agent.datasette.io/) y llm-coding-agent (https://github.com/simonw/llm-coding-agent).","Quizás la próxima versión del LLM integrará directamente el concepto de 'agente' en la biblioteca principal. Todavía estoy tratando de averiguar cómo sería eso.","Esta es la nueva versión de LLM, lanzada por Simon Willison el 4 de agosto de 2026: agrega soporte para seguimiento de razonamiento, respuestas de OpenAI, herramientas del lado del servidor y registro de eventos más inteligente: /2026/Aug/4/.","Parte de la serie de artículos 'Nuevas versiones de modelos de lenguaje grande (LLM)': /series/llm-releases/","Artículo anterior: MCP sin estado volvió a interesarme (y inspiró mcp-explorer y datasette-mcp): /2026/Jul/31/stateless-mcp/","Patrocíname con 10 dólares al mes y recibirás un resumen de correo seleccionado que recopila las noticias más importantes sobre LLM del mes."],"bodyTranslated":true,"bodySourceHash":"180cd773be407152","bodyTranslatedAt":"2026-08-08T12:22:32.425Z"},"fr":{"title":"LLM 0.32 publié : ajout de trajectoires d’inférence, d’OpenAI Responses, d’outils côté serveur, et de logs plus intelligents","summary":"Simon Willison a publié LLM 0.32, la version la plus importante depuis le lancement du projet. La nouvelle version prend en charge l’affichage des trajectoires d’inférence, des outils côté serveur, l’API OpenAI Responses, et utilise par défaut le modèle Luna GPT-5.6.","category":"Produits","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"LLM 0.32 publié : ajout de trajectoires d’inférence, d’OpenAI Responses, d’outils côté serveur, et de logs plus intelligents - Aioga Actualités IA","description":"Simon Willison a publié LLM 0.32, la version la plus importante depuis le lancement du projet. La nouvelle version prend en charge l’affichage des trajectoires d’inférence, des out...","url":"https://www.aioga.com/fr/news/cmsfdkkcs1ocxro2ed4z3s131/","contentTranslated":true,"sourceHash":"a2973b505f038783","translatedAt":"2026-08-05T01:22:53.728Z","articleBody":["J'ai publié ce matin LLM 0.32 : https://llm.datasette.io/en/stable/changelog.html#v0-32, il s'agit de la version la plus importante de LLM depuis le premier lancement du projet. La nouvelle version inclut le support des traces de raisonnement visibles, des outils pour les fournisseurs côté serveur, un journal SQLite à adressage du contenu repensé, de nouveaux modèles ainsi que de nouvelles fonctionnalités activées via l'API OpenAI Responses. J'ai également publié une nouvelle version du plugin llm-anthropic : https://github.com/simonw/llm-anthropic, qui comporte elle aussi de nombreuses mises à jour.","Exécuter des LLM sur un modèle de raisonnement affichera désormais leurs trajectoires de raisonnement sur la sortie d'erreur standard, vous permettant ainsi de voir leur contenu de « réflexion », alors que ces informations ne seront pas incluses dans la sortie standard que vous pourriez transmettre à d'autres outils. L'utilisation de -R/--hide-reasoning peut désactiver cette fonctionnalité.","Les LLM prennent en charge dès leur sortie la série de modèles GPT-5.6, et le nouveau modèle par défaut utilisant le \"prompt\" llm est désormais le GPT-5.6 Luna, bon marché mais puissant.","Les appels LLM peuvent désormais utiliser des outils côté serveur provenant de divers fournisseurs. OpenAI fournit un environnement d'exécution de code : https://llm.datasette.io/en/stable/openai-models.html#code-interpreter en tant qu'outil côté serveur ; les LLM peuvent maintenant exécuter des invites profitant de cet outil, comme suit :","OpenAI a également obtenu un outil WebSearch : https://llm.datasette.io/en/stable/openai-models.html#web-search.","llm-anthropic : https://github.com/simonw/llm-anthropic Le plugin a ajouté WebSearch : https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search, WebFetch : https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch, CodeExecution : https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution, ainsi que AnthropicMCP : https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector, cela semble être comme ceci :","Cela entraînera Anthropic à effectuer un appel MCP sur mon nouveau plugin datasette-mcp : https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp lors d'une seule interaction de requête/réponse avec son API.","La nouvelle commande LLM OpenAI Endpoint fournit un outil pour exécuter des invites sur n’importe quel point compatible OpenAI : https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it, une seule commande peut accomplir la tâche. Ces opérations ne sont pas enregistrées, ce qui en fait un outil pratique pouvant exécuter des invites ponctuelles sur n’importe quel objet en utilisant l’API LLM pour le langage universel.","Voici comment je l'utilise, en passant par uvx (sans installer LLM) et en combinant le plugin d'outils llm-tools-quickjs : https://github.com/simonw/llm-tools-quickjs, pour donner des instructions à Gemma 4 12B exécutée sur mon hôte local LM Studio : https://lmstudio.ai","L'API Python de LLM demandait auparavant de créer une session, puis d'envoyer un message à la fois. Il s'agit d'une abstraction des véritables caractéristiques de LLM, chaque requête transportant l'historique complet de tous les messages précédents. Pour certains cas plus avancés, cette abstraction commençait à devenir un obstacle, c'est pourquoi une nouvelle version a introduit le paramètre model.prompt(messages=[]), que l'on peut utiliser ainsi :","Avant, les LLM renvoyaient une séquence itérable de chaînes pour chaque invite. Cela fonctionnait bien lorsque le modèle renvoyait des réponses sous forme de chaîne, mais il était impossible de prédire les formes étranges que le modèle pourrait adopter. Aujourd'hui, de nombreux modèles renvoient un contenu mixte comprenant du texte de raisonnement, des chaînes de sortie, des appels d'outils, et même des pièces jointes d'images. Dans LLM 0.32, vous pouvez utiliser plutôt cette méthode : https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events","En combinant ces fonctionnalités, nous pouvons enfin fournir une implémentation robuste et semi-standard de l'API de complétion de chat OpenAI, que j'ai déjà publiée en tant que plugin llm-chat-completions-server : https://github.com/simonw/llm-chat-completions-server","Maintenant, vous pouvez utiliser la nouvelle commande endpoint llm openai via ce serveur pour exécuter des invites sur le LLM !","Le plus grand défi de ce type d'API réside dans la journalisation. Si nous voulons supporter un modèle où chaque requête ajoute une séquence de messages, il serait idéal d'éviter d'enregistrer tous les JSON répétés pour chaque session.","La solution est un nouveau stockage de messages adressable par le contenu : https://llm.datasette.io/en/stable/logging.html#the-message-store, modelé sur Git. Vous pouvez voir le nouveau schéma dans la documentation : https://llm.datasette.io/en/stable/logging.html#sql-schema, mais les commandes llm logs et llm logs --json ont été mises à jour pour pouvoir convertir ce format en une forme facile à utiliser.","Cette publication contient également plus de contenu. Notes de version de la 0.32 : https://llm.datasette.io/en/stable/changelog.html#v0-32 assez complètes, 0.32rc2 : https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30, 0.32rc : https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30, 0.32a3 : https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09, 0.32a2 : https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12 et 0.32a0 : https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 peuvent combler toute omission.","Les plugins LLM existants devraient tous pouvoir continuer à être utilisés, mais les plugins qui fournissent des modèles supplémentaires doivent être mis à jour vers la version 0.32 pour pouvoir participer pleinement au nouveau système d'événements en flux. La documentation contient un guide sur l'utilisation des messages structurés et des événements en flux pour implémenter les plugins : https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events.","J'ai mis à jour certains de mes propres plugins :","De nombreux changements apportés aux outils de base dans cette version ont été motivés par les besoins de Datasette Agent : https://agent.datasette.io/. Lorsque j'ai commencé à développer des LLM, le terme « agent » était très flou et je refusais de l'utiliser. En septembre 2025 : https://simonwillison.net/2025/Sep/18/agents/ J'ai adopté le point de vue selon lequel « les agents LLM atteignent leurs objectifs en exécutant des outils de manière itérative », ce qui était suffisamment mûr pour que je puisse complètement cesser d'éviter ce terme.","La chaîne d'outils peut maintenant être suspendue pour attendre l'approbation humaine : https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause, et reprendre à partir de l'historique des messages stockés : https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume — ces deux fonctionnalités sont nécessaires pour Datasette Agent.","Aujourd'hui, en regardant les LLM, ils commencent à sembler, à mes yeux, adopter une forme très similaire à celle des agents. Il y a un aspect très intéressant : vous pouvez utiliser un outil en ligne de commande pour combiner et assortir divers outils provenant de différentes sources et utilisant différents modèles, et tout cela peut se faire en une seule ligne de code. De plus, il comprend une bibliothèque Python suffisamment puissante pour construire des systèmes comme Datasette Agent (https://agent.datasette.io/) et llm-coding-agent (https://github.com/simonw/llm-coding-agent).","Peut-être que la prochaine version du LLM intégrera directement le concept d'« agent » dans la bibliothèque principale. J'essaie encore de comprendre à quoi cela ressemblerait.","Ceci est la nouvelle version du LLM, publiée par Simon Willison le 4 août 2026 : ajout du support pour le suivi du raisonnement, les réponses OpenAI, les outils côté serveur ainsi qu'un enregistrement des journaux plus intelligent : /2026/Aug/4/.","Une partie de la série d'articles « Nouvelles versions des grands modèles de langage (LLM) » : /series/llm-releases/","Article précédent : Le MCP sans état a ravivé mon intérêt (et a inspiré mcp-explorer et datasette-mcp) : /2026/Jul/31/stateless-mcp/","En me parrainant 10 dollars par mois, vous recevrez un résumé par e-mail sélectionné, regroupant les actualités les plus importantes sur les LLM du mois."],"bodyTranslated":true,"bodySourceHash":"180cd773be407152","bodyTranslatedAt":"2026-08-08T12:23:26.517Z"},"de":{"title":"LLM 0.32 veröffentlicht: hinzugefügte Inferenztrajektorien, OpenAI-Antworten, serverseitige Tools und intelligenteres Logging","summary":"Simon Willison veröffentlichte LLM 0.32, die wichtigste neue Version seit dem Start des Projekts. Die neue Version unterstützt die Darstellung von Inferenztrajektorien, serverseitigen Tools, die OpenAI Responses API und verwendet standardmäßig das GPT-5.6 Luna-Modell.","category":"产品更新","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"LLM 0.32 veröffentlicht: hinzugefügte Inferenztrajektorien, OpenAI-Antworten, serverseitige Tools und intelligenteres Logging - Aioga KI-News","description":"Simon Willison veröffentlichte LLM 0.32, die wichtigste neue Version seit dem Start des Projekts. Die neue Version unterstützt die Darstellung von Inferenztrajektorien, serverseiti...","url":"https://www.aioga.com/de/news/cmsfdkkcs1ocxro2ed4z3s131/","contentTranslated":true,"sourceHash":"a2973b505f038783","translatedAt":"2026-08-05T01:22:53.041Z","articleBody":["Ich habe heute Morgen LLM 0.32 veröffentlicht: https://llm.datasette.io/en/stable/changelog.html#v0-32, dies ist die bisher wichtigste neue Version von LLM seit der Erstveröffentlichung des Projekts. Die neue Version umfasst Unterstützung für sichtbare Reasoning-Pfade, Server-seitige Anbieter-Tools, ein neu gestaltetes Content-Addressed SQLite-Log, neue Modelle sowie neue Funktionen, die über die OpenAI Responses API aktiviert werden. Ich habe auch eine neue Version des llm-anthropic Plugins veröffentlicht: https://github.com/simonw/llm-anthropic, das ebenfalls viele Updates enthält.","Beim Ausführen von LLM auf einem Inferenzmodell werden deren Inferenzspuren jetzt auf die Standardfehlerausgabe angezeigt, sodass du ihren \"Denkprozess\" sehen kannst, während diese Informationen nicht in der Standardausgabe enthalten sind, die du möglicherweise an andere Werkzeuge weitergeben würdest. Die Verwendung von -R/--hide-reasoning kann diese Funktion deaktivieren.","LLM unterstützt direkt die GPT-5.6 Modellreihe, und das neue Standardmodell für die Verwendung von llm \"prompt\" ist jetzt das preiswerte, aber leistungsstarke GPT-5.6 Luna.","LLM-Aufrufe können jetzt serverseitige Werkzeuge von verschiedenen Anbietern verwenden. OpenAI bietet eine Code-Ausführungsumgebung an: https://llm.datasette.io/en/stable/openai-models.html#code-interpreter als serverseitiges Werkzeug; LLM kann jetzt Aufforderungen ausführen, die von der Nutzung dieses Werkzeugs profitieren, wie folgt:","OpenAI hat außerdem ein WebSearch-Tool erhalten: https://llm.datasette.io/en/stable/openai-models.html#web-search.","llm-anthropic: https://github.com/simonw/llm-anthropic Das Plugin hat WebSearch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search, WebFetch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch, CodeExecution: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution sowie AnthropicMCP: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector hinzugefügt, es sieht so aus:","Dies würde dazu führen, dass Anthropic in einer einzelnen Anfrage/Antwort-Interaktion mit ihrer API Aufrufe an mein neues datasette-mcp-Plugin https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp ausführt.","Der neue llm OpenAI-Endpunktbefehl bietet ein Tool, mit dem Sie Eingabeaufforderungen für jeden OpenAI-kompatiblen Endpunkt ausführen können: https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it, alles in einem einzigen Befehl. Diese Vorgänge werden nicht protokolliert, daher ist es ein praktisches Werkzeug, um einmalige Eingabeaufforderungen für jedes Objekt auszuführen, das die LLM API mit universeller Sprache verwendet.","Dies ist, wie ich es benutze, über uvx (kein LLM erforderlich) und in Kombination mit dem llm-tools-quickjs: https://github.com/simonw/llm-tools-quickjs Tool-Plugin, um das auf meinem lokalen Host LM Studio: https://lmstudio.ai laufende Gemma 4 12B zu prompten:","Die Python-API von LLM verlangte früher, dass man eine Sitzung erstellt und dann jeweils eine Nachricht sendet. Dies ist eine Abstraktion der eigentlichen Eigenschaften von LLM, wobei jede Anfrage den vollständigen Verlauf aller vorherigen Nachrichten enthält. Für einige fortgeschrittenere Anwendungsfälle beginnt diese Abstraktion jedoch zu einem Hindernis zu werden, daher wurde in der neuen Version der Parameter model.prompt(messages=[]) eingeführt, der folgendermaßen verwendet werden kann:","Früher gab LLM von jedem Prompt aus eine iterable Sequenz von Strings zurück. Dies funktionierte gut, wenn das Modell String-Antworten zurückgab, war aber unvorhersehbar, wenn das Modell seltsame Formen entwickelte. Heutzutage geben viele Modelle eine Mischung aus Begründungstexten, Ausgabestrings, Werkzeugaufrufen und sogar Bildanhängen zurück. In LLM 0.32 kannst du stattdessen diese Methode verwenden: https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events","In Verbindung mit diesen Funktionen können wir schließlich eine robuste halbstandardisierte OpenAI-Chat-Abschluss-API-Implementierung anbieten. Ich habe sie bereits als llm-chat-completions-server veröffentlicht: https://github.com/simonw/llm-chat-completions-server","Jetzt kannst du über diesen Server den neuen LLM OpenAI Endpoint-Befehl verwenden, um Hinweise für das LLM auszuführen!","Die größere Herausforderung bei diesem Typ von API liegt im Logging. Wenn wir das Muster unterstützen wollen, bei dem bei jeder Anfrage eine Nachrichtensequenz angehängt wird, sollte es idealerweise vermieden werden, jedes Mal für jede Sitzung alle wiederholten JSONs zu protokollieren.","Die Lösung ist ein neues inhaltsadressierbares Nachrichtenspeicher: https://llm.datasette.io/en/stable/logging.html#the-message-store, modelliert nach Git. Du kannst das neue Schema in der Dokumentation sehen: https://llm.datasette.io/en/stable/logging.html#sql-schema, aber die Befehle llm logs und llm logs --json wurden beide aktualisiert und können dieses Format wieder in eine leicht verwendbare Form umwandeln.","Dieses Release enthält noch mehr Inhalte. Release Notes für 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32 sind ziemlich umfassend, 0.32rc2: https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30, 0.32rc: https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30, 0.32a3: https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09, 0.32a2: https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12 und 0.32a0: https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 können alle fehlenden Informationen ergänzen.","Vorhandene LLM-Plugins sollten weiterhin verwendet werden können, aber Plugins, die zusätzliche Modelle bereitstellen, müssen auf Version 0.32 aktualisiert werden, um vollständig am neuen Stream-Event-System teilzunehmen. In der Dokumentation gibt es eine Anleitung zur Implementierung von Plugins mit strukturierten Nachrichten und Stream-Events: https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events.","Ich habe einige meiner eigenen Plugins aktualisiert:","Viele der in dieser Version veröffentlichten Änderungen an den grundlegenden Werkzeugen wurden durch die Bedürfnisse des Datasette Agent: https://agent.datasette.io/ vorangetrieben. Als ich begann, LLM zu entwickeln, war der Begriff „Agent“ sehr unklar definiert, und ich weigerte mich damals, ihn zu verwenden. Bis September 2025: https://simonwillison.net/2025/Sep/18/agents/ habe ich die Ansicht übernommen, dass „LLM-Agenten Werkzeuge in einem Kreislauf ausführen, um Ziele zu erreichen“, was bereits ausgereift genug war, dass ich ganz aufhören konnte, den Begriff zu vermeiden.","Die Toolchain kann jetzt pausiert werden, um auf manuelle Genehmigung zu warten: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause, und aus dem gespeicherten Nachrichtenverlauf wieder aufgenommen werden: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume – beides wird vom Datasette Agent benötigt.","Heute habe ich mir LLM angesehen, und es beginnt für mich, eine sehr agentenähnliche Form anzunehmen. Ein sehr interessantes Merkmal ist, dass man mit einem Kommandozeilentool verschiedene Werkzeuge aus unterschiedlichen Quellen und mit verschiedenen Modellen kombinieren kann – alles mit einer einzigen Codezeile. Außerdem enthält es eine leistungsstarke Python-Bibliothek, mit der man Systeme wie Datasette Agent (https://agent.datasette.io/) und llm-coding-agent (https://github.com/simonw/llm-coding-agent) erstellen kann.","Vielleicht wird die nächste Version des LLM das Konzept des „Agenten“ direkt in die Kernbibliothek integrieren. Ich versuche immer noch herauszufinden, wie das aussehen wird.","Dies ist die neue Version von LLM, veröffentlicht von Simon Willison am 4. August 2026: Hinzugekommen ist die Unterstützung für Nachverfolgung von Schlussfolgerungen, OpenAI-Antworten, serverseitige Werkzeuge sowie intelligentere Protokollierung: /2026/Aug/4/.","Teil der Artikelserie „Neue Versionen großer Sprachmodelle (LLM)“: /series/llm-releases/","Vorheriger Artikel: Stateless MCP hat mein Interesse wieder geweckt (und inspirierte mcp-explorer und datasette-mcp): /2026/Jul/31/stateless-mcp/","Sponsern Sie mich jeden Monat mit 10 US-Dollar, und Sie erhalten eine ausgewählte E-Mail-Zusammenfassung, die die wichtigsten LLM-Entwicklungen des Monats zusammenfasst."],"bodyTranslated":true,"bodySourceHash":"180cd773be407152","bodyTranslatedAt":"2026-08-08T12:23:19.601Z"},"pt-BR":{"title":"Lançamento do LLM 0.32: adição de trajetórias de inferência, respostas OpenAI, ferramentas do lado do servidor e logging mais inteligente","summary":"Simon Willison lançou o LLM 0.32, a versão nova mais importante desde o lançamento do projeto. A nova versão suporta a exibição de trajetórias de inferência, ferramentas do lado do servidor, a API OpenAI Responses e utiliza o modelo Luna do GPT-5.6 por padrão.","category":"产品更新","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Lançamento do LLM 0.32: adição de trajetórias de inferência, respostas OpenAI, ferramentas do lado do servidor e logging mais inteligente - Aioga Notícias de IA","description":"Simon Willison lançou o LLM 0.32, a versão nova mais importante desde o lançamento do projeto. A nova versão suporta a exibição de trajetórias de inferência, ferramentas do lado do...","url":"https://www.aioga.com/pt-BR/news/cmsfdkkcs1ocxro2ed4z3s131/","contentTranslated":true,"sourceHash":"a2973b505f038783","translatedAt":"2026-08-05T01:23:02.536Z","articleBody":["Hoje de manhã lancei o LLM 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32, que é a versão mais importante do LLM desde o lançamento inicial do projeto. A nova versão inclui suporte a rastros de raciocínio visíveis, ferramentas de provedores no lado do servidor, registro SQLite com endereçamento de conteúdo redesenhado, novos modelos e novos recursos habilitados através da OpenAI Responses API. Também lancei uma nova versão do plugin llm-anthropic: https://github.com/simonw/llm-anthropic, que também tem muitas atualizações.","Executar LLMs em modelos de raciocínio agora exibirá suas trajetórias de raciocínio na saída de erro padrão, para que você possa ver o conteúdo do seu “pensamento”, e essas informações não estarão incluídas na saída padrão que você pode passar para outras ferramentas. Usar -R/--hide-reasoning pode desativar esse recurso.","Os LLMs suportam prontamente a série de modelos GPT-5.6 e o novo modelo padrão usando \"prompt\" do llm agora é o GPT-5.6 Luna, barato mas poderoso.","As chamadas de LLM agora podem usar ferramentas do lado do servidor de vários provedores. A OpenAI oferece um ambiente de execução de código: https://llm.datasette.io/en/stable/openai-models.html#code-interpreter como uma ferramenta do lado do servidor; LLM agora pode executar prompts que se beneficiam do uso dessa ferramenta, conforme mostrado a seguir:","A OpenAI também adquiriu uma ferramenta WebSearch: https://llm.datasette.io/en/stable/openai-models.html#web-search.","llm-anthropic: https://github.com/simonw/llm-anthropic O plugin adicionou WebSearch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search, WebFetch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch, CodeExecution: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution, e AnthropicMCP: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector, parece ser assim:","Isso fará com que a Anthropic execute chamadas MCP no meu novo plugin datasette-mcp: https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp em interações únicas de solicitação/resposta com sua API.","O novo comando de endpoint llm do OpenAI fornece uma ferramenta que permite executar prompts em qualquer endpoint compatível com OpenAI: https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it, tudo com um único comando. Essas operações não são registradas, portanto, é uma ferramenta conveniente para executar prompts únicos em qualquer objeto que use a linguagem universal do mundo da API LLM.","Esta é a maneira como eu uso isso, através do uvx (sem necessidade de instalar o LLM) e combinando com o plugin llm-tools-quickjs: https://github.com/simonw/llm-tools-quickjs, para enviar prompts ao Gemma 4 12B que está rodando no meu host local LM Studio: https://lmstudio.ai","A API do Python do LLM costumava exigir que você criasse uma sessão e, em seguida, enviasse uma mensagem de cada vez. Esta é uma abstração das verdadeiras características do LLM, pois cada solicitação carrega o histórico completo de todas as mensagens anteriores. Para alguns casos mais avançados, essa abstração começava a se tornar um obstáculo, portanto, a nova versão introduziu o parâmetro model.prompt(messages=[]), que pode ser usado assim:","Antes, o LLM retornava uma sequência iterável de strings a partir de cada prompt. Isso funcionava bem quando o modelo retornava respostas em forma de string, mas era imprevisível em relação a formas estranhas que o modelo poderia gerar. Hoje, muitos modelos retornam um conteúdo misto de texto de raciocínio, strings de saída, chamadas de ferramentas, e até anexos de imagens. No LLM 0.32, você pode usar este método: https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events","Combinando essas funções, finalmente podemos fornecer uma implementação robusta de API de conclusão de chat semi-padrão da OpenAI, que eu já publiquei como o plugin llm-chat-completions-server: https://github.com/simonw/llm-chat-completions-server","Agora, você pode usar o novo comando de endpoint llm openai através deste servidor para executar prompts no LLM!","O maior desafio desse tipo de API está na gravação de logs. Se quisermos suportar o padrão de adicionar sequências de mensagens a cada requisição, o ideal seria evitar registrar todo o JSON repetido a cada sessão.","A solução é um novo armazenamento de mensagens endereçável por conteúdo: https://llm.datasette.io/en/stable/logging.html#the-message-store, modelando-se no Git. Você pode ver o novo esquema na documentação: https://llm.datasette.io/en/stable/logging.html#sql-schema, mas os comandos llm logs e llm logs --json foram atualizados e podem converter esse formato de volta para uma forma de fácil uso.","Esta versão traz ainda mais conteúdos. Notas da versão 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32 são bastante completas, 0.32rc2: https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30, 0.32rc: https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30, 0.32a3: https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09, 0.32a2: https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12 e 0.32a0: https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 podem preencher quaisquer lacunas.","Os plugins LLM existentes devem continuar funcionando, mas os plugins que fornecem modelos adicionais precisam ser atualizados para a versão 0.32 para poder participar totalmente do novo sistema de eventos de fluxo. A documentação fornece um guia sobre como implementar plugins usando mensagens estruturadas e eventos de fluxo: https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events.","Eu atualizei alguns dos meus próprios plugins:","Muitas das alterações nas ferramentas de baixo nível lançadas desta vez são impulsionadas pelas necessidades do Datasette Agent: https://agent.datasette.io/. Quando comecei a desenvolver LLMs, o termo “agente” tinha uma definição muito vaga e, na época, eu me recusei a usá-lo. Até setembro de 2025: https://simonwillison.net/2025/Sep/18/agents/ aceitei a visão de que “agentes LLM executam ferramentas em loop para atingir objetivos”, que já estava suficientemente madura, e eu poderia parar totalmente de evitar o termo.","A cadeia de ferramentas agora pode ser pausada para aguardar aprovação humana: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause, e ser retomada a partir do histórico de mensagens armazenado: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume — ambos são necessários para o Datasette Agent.","Hoje, olhando para o LLM, ele começou a se mostrar, aos meus olhos, com uma forma muito parecida com a de um agente. Um ponto muito interessante é que você pode usar uma ferramenta de linha de comando para misturar diversas ferramentas de diferentes fontes e usando modelos diferentes, resolvendo tudo com uma linha de código, e ainda inclui uma biblioteca Python suficientemente poderosa para construir sistemas como o Datasette Agent (https://agent.datasette.io/) e o llm-coding-agent (https://github.com/simonw/llm-coding-agent).","Talvez a próxima versão do LLM integre diretamente o conceito de 'agente' na biblioteca principal. Eu ainda estou tentando entender como seria isso.","Esta é a nova versão do LLM, lançada por Simon Willison em 4 de agosto de 2026: adicionou suporte para rastreamento de raciocínio, respostas da OpenAI, ferramentas do lado do servidor e registro de logs mais inteligente: /2026/Aug/4/.","Parte da série de artigos \"Novas versões de grandes modelos de linguagem (LLM) lançadas\": /series/llm-releases/","Post anterior: O MCP sem estado reacendeu meu interesse (e inspirou o mcp-explorer e o datasette-mcp): /2026/Jul/31/stateless-mcp/","Patrocine-me com 10 dólares por mês e receba um resumo de e-mails selecionados, reunindo as atualizações mais importantes de LLM do mês."],"bodyTranslated":true,"bodySourceHash":"180cd773be407152","bodyTranslatedAt":"2026-08-08T12:24:06.822Z"},"ru":{"title":"Вышел LLM 0.32: добавлены траектории вывода, ответы OpenAI, серверные инструменты и более умное логирование","summary":"Саймон Уиллисон выпустил LLM 0.32 — самую важную новую версию с момента запуска проекта. Новая версия поддерживает отображение траекторий вывода, серверные инструменты, API OpenAI Responses и по умолчанию использует модель GPT-5.6 Luna.","category":"产品更新","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Вышел LLM 0.32: добавлены траектории вывода, ответы OpenAI, серверные инструменты и более умное логирование - Aioga Новости ИИ","description":"Саймон Уиллисон выпустил LLM 0.32 — самую важную новую версию с момента запуска проекта. Новая версия поддерживает отображение траекторий вывода, серверные инструменты, API OpenAI...","url":"https://www.aioga.com/ru/news/cmsfdkkcs1ocxro2ed4z3s131/","contentTranslated":true,"sourceHash":"a2973b505f038783","translatedAt":"2026-08-05T01:23:01.510Z","articleBody":["Сегодня утром я выпустил LLM 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32, это самая важная новая версия LLM с момента первого выпуска проекта. Новая версия включает поддержку видимых следов рассуждений, инструменты серверных провайдеров, переработанный журнал SQLite с адресацией по содержимому, новые модели и новые функции, включенные через OpenAI Responses API. Я также выпустил новую версию плагина llm-anthropic: https://github.com/simonw/llm-anthropic, которая также получила множество обновлений.","Запуск LLM на модели рассуждений теперь будет выводить их траектории рассуждений в стандартный поток ошибок, так что вы можете видеть их «мышление», и эта информация не будет включена в стандартный вывод, который вы можете передавать другим инструментам. Использование -R/--hide-reasoning позволяет отключить эту функцию.","LLM из коробки поддерживает серию моделей GPT-5.6, и новая модель по умолчанию для использования \"prompt\" в llm теперь является дешёвым, но мощным GPT-5.6 Luna.","Вызов LLM теперь может использовать серверные инструменты от различных поставщиков. OpenAI предоставляет среду выполнения кода: https://llm.datasette.io/en/stable/openai-models.html#code-interpreter в качестве серверного инструмента; теперь LLM может выполнять подсказки, которые выигрывают от использования этого инструмента, как показано ниже:","OpenAI также получил инструмент WebSearch: https://llm.datasette.io/en/stable/openai-models.html#web-search.","llm-anthropic: https://github.com/simonw/llm-anthropic Плагин добавил WebSearch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search, WebFetch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch, CodeExecution: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution, а также AnthropicMCP: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector, выглядит это так:","Это приведет к тому, что Anthropic выполняет вызовы MCP для моего нового плагина datasette-mcp: https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp в однократных запросах/ответах при взаимодействии с его API.","Новая команда llm openai endpoint предоставляет инструмент, который позволяет выполнять подсказки на любом совместимом с OpenAI конце: https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it — всё выполняется одной командой. Эти операции не записываются, поэтому это удобный инструмент для одноразового выполнения подсказок на любых объектах, использующих универсальный язык мира LLM API.","Вот как я использую это, через uvx (без необходимости установки LLM) и в сочетании с плагином инструментов llm-tools-quickjs: https://github.com/simonw/llm-tools-quickjs, для подсказок Gemma 4 12B, работающей на моем локальном хосте LM Studio: https://lmstudio.ai","Python API LLM раньше требовал, чтобы вы создавали сеанс и отправляли одно сообщение за раз. Это абстракция настоящих возможностей LLM, каждый запрос несет в себе полную историю всех предыдущих сообщений. Для некоторых более сложных случаев эта абстракция начинает становиться препятствием, поэтому в новой версии введён параметр model.prompt(messages=[]), который можно использовать следующим образом:","Ранее LLM возвращал итерабельную последовательность строк из каждого запроса. Это хорошо работало, когда модель возвращала строковые ответы, но невозможно было предсказать странные формы, которые могла бы принять модель. В настоящее время многие модели возвращают смесь рассуждающего текста, строковых выходных данных, вызовов инструментов и даже графических вложений. В LLM 0.32 вы можете вместо этого использовать этот метод: https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events","Объединив эти функции, мы наконец-то можем предоставить надёжную полуофициальную реализацию API завершения чата OpenAI. Я уже опубликовал её как плагин llm-chat-completions-server: https://github.com/simonw/llm-chat-completions-server","Теперь вы можете использовать новый конечный пункт llm openai через этот сервер для запуска подсказок LLM!","Большая проблема для такого типа API заключается в ведении журналов. Если мы хотим поддерживать режим, при котором каждый запрос добавляется к последовательности сообщений, в идеале можно избежать записи всех повторяющихся JSON для каждой сессии.","Решение — это новое хранилище сообщений с адресацией по содержимому: https://llm.datasette.io/en/stable/logging.html#the-message-store, моделируемое по примеру Git. Вы можете увидеть новую схему в документации: https://llm.datasette.io/en/stable/logging.html#sql-schema, но команды llm logs и llm logs --json были обновлены и могут преобразовывать этот формат обратно в удобную для использования форму.","В этом выпуске есть ещё больше материалов. Примечания к выпуску 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32 достаточно полные, 0.32rc2: https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30, 0.32rc: https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30, 0.32a3: https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09, 0.32a2: https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12 и 0.32a0: https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 могут заполнить любые пропуски.","Существующие плагины LLM должны продолжать работать, но плагины, предоставляющие дополнительные модели, необходимо обновить до версии 0.32, чтобы полностью участвовать в новой системе потоковых событий. В документации есть руководство по использованию структурированных сообщений и потоковых событий для реализации плагинов: https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events.","Я обновил некоторые из моих собственных плагинов:","Многие изменения в нижележащих инструментах в этом выпуске были продиктованы потребностями Datasette Agent: https://agent.datasette.io/. Когда я начал разрабатывать LLM, термин «агент» был очень неопределённым, поэтому я тогда отказывался его использовать. К сентябрю 2025 года: https://simonwillison.net/2025/Sep/18/agents/ я принял следующую точку зрения: «LLM-агенты достигают цели, последовательно используя инструменты», что было достаточно зрелым, чтобы я полностью перестал избегать этого термина.","Инструментальная цепочка теперь может приостанавливаться для ожидания ручного одобрения: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause и восстанавливаться из сохранённой истории сообщений: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume — оба этих момента необходимы агенту Datasette.","Сегодня, когда я смотрю на LLM, он начинает представляться мне в виде, очень похожем на агента (agent). Один очень интересный момент заключается в том, что вы можете использовать инструмент командной строки для смешивания различных инструментов из разных источников и с использованием разных моделей — это можно сделать одной строкой кода, и при этом он включает достаточно мощную библиотеку Python, с помощью которой можно создавать системы, такие как Datasette Agent (https://agent.datasette.io/) и llm-coding-agent (https://github.com/simonw/llm-coding-agent).","Возможно, следующая версия LLM напрямую интегрирует концепцию «агента» в основную библиотеку. Я всё ещё пытаюсь понять, как это будет выглядеть.","Это новая версия LLM, выпущенная Саймоном Виллисоном 4 августа 2026 года: добавлена поддержка отслеживания рассуждений, откликов OpenAI, серверных инструментов и более умной записи логов: /2026/Aug/4/.","Часть серии статей «Новые версии больших языковых моделей (LLM)»: /series/llm-releases/","Предыдущая статья: Безсостояночный MCP снова вызвал мой интерес (и вдохновил mcp-explorer и datasette-mcp): /2026/Jul/31/stateless-mcp/","Спонсируйте меня на 10 долларов в месяц, и вы будете получать избранные email-резюме с самыми важными событиями месяца в области LLM."],"bodyTranslated":true,"bodySourceHash":"180cd773be407152","bodyTranslatedAt":"2026-08-08T12:24:10.811Z"},"ar":{"title":"إصدار LLM 0.32: إضافة مسارات استدلال، واستجابات OpenAI، وأدوات على جانب الخادم، وتسجيل أكثر ذكاء","summary":"أصدر سايمون ويليسون إصدار LLM 0.32، وهو أهم إصدار جديد منذ إطلاق المشروع. يدعم الإصدار الجديد عرض مسارات الاستدلال، وأدوات على جانب الخادم، وواجهة برمجة تطبيقات OpenAI Responses، ويستخدم نموذج GPT-5.6 Luna بشكل افتراضي.","category":"产品更新","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"إصدار LLM 0.32: إضافة مسارات استدلال، واستجابات OpenAI، وأدوات على جانب الخادم، وتسجيل أكثر ذكاء - Aioga أخبار الذكاء الاصطناعي","description":"أصدر سايمون ويليسون إصدار LLM 0.32، وهو أهم إصدار جديد منذ إطلاق المشروع. يدعم الإصدار الجديد عرض مسارات الاستدلال، وأدوات على جانب الخادم، وواجهة برمجة تطبيقات OpenAI Responses، و...","url":"https://www.aioga.com/ar/news/cmsfdkkcs1ocxro2ed4z3s131/","contentTranslated":true,"sourceHash":"a2973b505f038783","translatedAt":"2026-08-05T01:23:11.268Z","articleBody":["لقد أصدرت هذا الصباح LLM 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32، وهذا هو الإصدار الجديد الأكثر أهمية من LLM منذ إطلاق المشروع لأول مرة. يتضمن الإصدار الجديد دعم مسار الاستدلال المرئي، أدوات مزود الخادم، سجلات SQLite المعاد تصميمها للتوجيه بالمحتوى، نماذج جديدة، ووظائف جديدة مفعّلة من خلال OpenAI Responses API. كما أصدرت أيضًا إصدارًا جديدًا من ملحق llm-anthropic: https://github.com/simonw/llm-anthropic، والذي يحتوي أيضًا على العديد من التحديثات.","تشغيل LLM على نموذج الاستدلال الآن سيعرض مسارات استدلالها على مخرجات الخطأ القياسية، بحيث يمكنك رؤية محتوى 'تفكيرها'، وهذه المعلومات لن تكون مضمنة في المخرجات القياسية التي قد تمررها إلى أدوات أخرى. يمكن استخدام -R/--hide-reasoning لإيقاف هذه الميزة.","يدعم LLM مجموعة نماذج GPT-5.6 مباشرة دون إعداد، والنموذج الافتراضي الجديد لاستخدام \"prompt\" في llm الآن هو GPT-5.6 Luna الرخيص ولكنه قوي الأداء.","يمكن الآن استخدام استدعاءات LLM أدوات جانب الخادم من موفِّرين مختلفين. تقدم OpenAI بيئة تنفيذ الشيفرة: https://llm.datasette.io/en/stable/openai-models.html#code-interpreter كأداة جانب الخادم؛ يمكن لـ LLM الآن تشغيل المطالبات التي تستفيد من هذه الأداة كما يلي:","حصلت OpenAI أيضًا على أداة WebSearch: https://llm.datasette.io/en/stable/openai-models.html#web-search","llm-anthropic: https://github.com/simonw/llm-anthropic المكوّن الإضافي أضاف WebSearch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search، WebFetch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch، CodeExecution: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution، وكذلك AnthropicMCP: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector، ويبدو أنه هكذا:","هذا سيؤدي إلى أن يقوم Anthropic باستدعاء MCP للإضافة الجديدة الخاصة بي datasette-mcp: https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp أثناء التفاعل بالطلب/الاستجابة الفردي مع واجهة برمجة التطبيقات الخاصة به.","توفر أوامر نقطة النهاية الجديدة llm من OpenAI أداة يمكنها تنفيذ أي موجه على أي نقطة نهاية متوافقة مع OpenAI: https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it، يمكن إكمال ذلك بأمر واحد فقط. لن يتم تسجيل هذه العمليات، لذلك فهي أداة مناسبة لتشغيل موجهات لمرة واحدة على أي كائن يستخدم لغة العالم العامة API الخاصة بـ LLM.","هذه هي الطريقة التي أستخدمها، عبر uvx (دون الحاجة لتثبيت LLM) وبالاقتران مع أداة llm-tools-quickjs: https://github.com/simonw/llm-tools-quickjs، لإعطاء التعليمات لـ Gemma 4 12B التي تعمل على جهاز الاستوديو المحلي LM Studio: https://lmstudio.ai.","كان واجهة برمجة تطبيقات Python الخاصة بـ LLM تتطلب منك سابقًا إنشاء جلسة، ثم إرسال رسالة واحدة في كل مرة. هذا هو تجريد للخصائص الحقيقية لـ LLM، حيث يحمل كل طلب كامل سجل جميع الرسائل السابقة. بالنسبة لبعض الحالات الأكثر تقدمًا، يبدأ هذا التجريد في أن يصبح عقبة، لذلك قدمت النسخة الجديدة معامل model.prompt(messages=[])، ويمكن استخدامه على هذا النحو:","في السابق، كان LLM يُرجع سلسلة قابلة للتكرار من كل موجه. كان هذا يعمل بشكل جيد عند أن يُرجع النموذج استجابة نصية، لكن كان من الصعب التنبؤ بالشكل الغريب الذي قد يتطور إليه النموذج. في الوقت الحاضر، تُرجع العديد من النماذج مزيجًا من نص الاستدلال، نص الإخراج، استدعاءات الأدوات، وحتى مرفقات الصور. في LLM 0.32، يمكنك استخدام هذه الطريقة بدلاً من ذلك: https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events","بدمج هذه الميزات، يمكننا أخيراً تقديم تنفيذ API لاستكمالات الدردشة من OpenAI شبه القياسية والثابتة، لقد قمت بنشره كمكوّن إضافي llm-chat-completions-server: https://github.com/simonw/llm-chat-completions-server","الآن، يمكنك استخدام خادمك لأمر نقطة نهاية llm openai الجديدة لتشغيل المطالبات على LLM!","التحدي الأكبر لهذا النوع من واجهات برمجة التطبيقات يكمن في تسجيل السجلات. إذا أردنا دعم نمط إضافة تسلسل الرسائل في كل طلب، من المثالي تجنب تسجيل كل JSON مكرر لكل جلسة.","الحل هو تخزين الرسائل الجديد القابل للعناوين: https://llm.datasette.io/en/stable/logging.html#the-message-store، يحاكي نموذج Git. يمكنك الاطلاع على النموذج الجديد في الوثائق: https://llm.datasette.io/en/stable/logging.html#sql-schema، لكن كل من أوامر llm logs و llm logs --json قد تم ترقيتها، ويمكنها تحويل هذا التنسيق مرة أخرى إلى شكل سهل الاستخدام.","هذا الإصدار يحتوي على المزيد من المحتوى. ملاحظات الإصدار 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32 شاملة تمامًا، و0.32rc2: https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30، و0.32rc: https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30، و0.32a3: https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09، و0.32a2: https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12 و0.32a0: https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 يمكن أن تكمل أي جزء مفقود.","يمكن الاستمرار في استخدام كافة ملحقات LLM الحالية، ولكن الملحقات التي توفر نماذج إضافية تحتاج إلى التحديث إلى الإصدار 0.32 للمشاركة الكاملة في نظام أحداث التدفق الجديد. تحتوي الوثائق على دليل حول استخدام الرسائل المهيكلة وأحداث التدفق لتنفيذ الملحقات: https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events.","لقد قمت بتحديث بعض الإضافات الخاصة بي:","تم دفع العديد من التغييرات على الأدوات الأساسية التي صدرت هذه المرة بواسطة احتياجات Datasette Agent: https://agent.datasette.io/. عندما بدأت تطوير LLM، كان تعريف مصطلح 'الوكيل' غامضًا جدًا، ولذلك رفضت استخدامه في ذلك الوقت. بحلول سبتمبر 2025: https://simonwillison.net/2025/Sep/18/agents/ قبلت وجهة النظر التالية: 'يعمل وكلاء LLM عن طريق تشغيل الأدوات بشكل دوري لتحقيق الهدف'، وكان هذا كافيًا للنضوج بحيث يمكنني التوقف تمامًا عن تجنب استخدام هذا المصطلح.","يمكن الآن للأداة إيقاف التشغيل مؤقتًا في انتظار الموافقة اليدوية: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause، واستئنافها من سجل الرسائل المخزّن: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume — كلا الأمرين مطلوبان من قبل وكيل Datasette.","اليوم شاهدت LLM، وبدأ يظهر لي بشكل يشبه الوكيل (agent) للغاية. هناك شيء رائع، وهو أنه يمكنك باستخدام أداة سطر الأوامر دمج أدوات مختلفة من مصادر مختلفة وتستخدم نماذج مختلفة، ويمكنك إنجاز ذلك بسطر كود واحد، كما أنه يتضمن مكتبة Python قوية بما يكفي لبناء أنظمة مثل Datasette Agent (https://agent.datasette.io/) و llm-coding-agent (https://github.com/simonw/llm-coding-agent).","ربما الإصدار القادم من LLM سيدمج مفهوم \"الوكيل\" مباشرة في المكتبة الأساسية. ما زلت أحاول معرفة كيف سيكون ذلك.","هذا هو الإصدار الجديد من LLM، أطلقه سايمون ويليسون في 4 أغسطس 2026: يتضمن دعمًا جديدًا لتتبع الاستدلال، واستجابات OpenAI، وأدوات الخادم، بالإضافة إلى تسجيل أكثر ذكاءً للسجلات: /2026/Aug/4/.","جزء من سلسلة المقالات \"إصدارات النسخ الجديدة لنماذج اللغة الكبيرة (LLM)\": /series/llm-releases/","المقال السابق: أعاد MCP غير الحالة إثارة اهتمامي (وألهم mcp-explorer و datasette-mcp): /2026/Jul/31/stateless-mcp/","ادعموني بمبلغ 10 دولارات شهريًا للحصول على ملخص بريد إلكتروني مختار يجمع أهم الأخبار عن نماذج اللغة الكبيرة (LLM) خلال الشهر."],"bodyTranslated":true,"bodySourceHash":"180cd773be407152","bodyTranslatedAt":"2026-08-08T12:25:06.498Z"},"hi":{"title":"एलएलएम 0.32 जारी: अतिरिक्त अनुमान प्रक्षेपवक्र, ओपनएआई प्रतिक्रियाएं, सर्वर-साइड टूल और स्मार्ट लॉगिंग","summary":"साइमन विलिसन ने एलएलएम 0.32 जारी किया, जो परियोजना के लॉन्च के बाद से सबसे महत्वपूर्ण नया संस्करण है। नया संस्करण अनुमान प्रक्षेपवक्र, सर्वर-साइड टूल, OpenAI रिस्पॉन्स API प्रदर्शित करने का समर्थन करता है, और डिफ़ॉल्ट रूप से GPT-5.6 लूना मॉडल का उपयोग करता है।","category":"产品更新","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"एलएलएम 0.32 जारी: अतिरिक्त अनुमान प्रक्षेपवक्र, ओपनएआई प्रतिक्रियाएं, सर्वर-साइड टूल और स्मार्ट लॉगिंग - Aioga AI समाचार","description":"साइमन विलिसन ने एलएलएम 0.32 जारी किया, जो परियोजना के लॉन्च के बाद से सबसे महत्वपूर्ण नया संस्करण है। नया संस्करण अनुमान प्रक्षेपवक्र, सर्वर-साइड टूल, OpenAI रिस्पॉन्स API प्रदर्शि...","url":"https://www.aioga.com/hi/news/cmsfdkkcs1ocxro2ed4z3s131/","contentTranslated":true,"sourceHash":"a2973b505f038783","translatedAt":"2026-08-05T01:23:11.377Z","articleBody":["मैंने आज सुबह LLM 0.32 जारी किया: https://llm.datasette.io/en/stable/changelog.html#v0-32, यह LLM का परियोजना की प्रारंभिक रिलीज़ के बाद सबसे महत्वपूर्ण नया संस्करण है। नए संस्करण में दृश्य तर्क पथ के समर्थन, सर्वर-साइड प्रदाता उपकरण, पुन: डिज़ाइन किए गए सामग्री-आधारित SQLite लॉग, नए मॉडल और OpenAI Responses API के माध्यम से सक्षम किए गए नए फीचर्स शामिल हैं। मैंने llm-anthropic प्लगइन का नया संस्करण भी जारी किया: https://github.com/simonw/llm-anthropic, इसमें भी भारी अपडेट शामिल हैं।","अब在推理模型上运行 LLM会显示它们的推理轨迹到标准错误输出，因此你可以看到它们的“思考”内容，而这些信息不会包含在你可能传给其他工具的标准输出中。使用 -R/--hide-reasoning 可以关闭此功能。","LLM बॉक्स से बाहर ही GPT-5.6 मॉडल श्रृंखला का समर्थन करता है, और llm \"prompt\" का नया डिफ़ॉल्ट मॉडल अब सस्ता लेकिन शक्तिशाली GPT-5.6 Luna है।","LLM कॉल अब विभिन्न प्रदाताओं से सर्वर-साइड टूल्स का उपयोग कर सकते हैं। OpenAI एक कोड निष्पादन वातावरण प्रदान करता है: https://llm.datasette.io/en/stable/openai-models.html#code-interpreter एक सर्वर-साइड टूल के रूप में; LLM अब इस टूल का लाभ उठाने वाले प्रॉम्प्ट चला सकते हैं, जैसा कि नीचे दिखाया गया है:","OpenAI ने एक WebSearch भी प्राप्त किया: https://llm.datasette.io/en/stable/openai-models.html#web-search उपकरण।","llm-anthropic: https://github.com/simonw/llm-anthropic प्लगइन में WebSearch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search, WebFetch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch, CodeExecution: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution, और AnthropicMCP: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector शामिल किए गए हैं, यह इस तरह दिखता है:","इससे Anthropic मेरे नए datasette-mcp: https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp प्लगइन पर MCP कॉल निष्पादित करेगा जब वह इसके API के साथ एकल अनुरोध/प्रतिक्रिया इंटरैक्शन में होता है।","नया llm openai endpoint कमांड एक टूल प्रदान करता है, जो किसी भी OpenAI-संगत एंडपॉइंट पर संकेत चला सकता है: https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it, केवल एक कमांड लाइन में। ये ऑपरेशन्स रिकॉर्ड नहीं किए जाते हैं, इसलिए यह एक सुविधाजनक टूल है, जो किसी भी ऑब्जेक्ट पर LLM API विश्वव्यापी भाषा का एक बार उपयोग करने वाला संकेत चला सकता है।","यह है कि मैं इसे कैसे उपयोग करता हूँ, uvx (LLM इंस्टॉल किए बिना) और llm-tools-quickjs: https://github.com/simonw/llm-tools-quickjs टूल प्लगइन का संयोजन करके, जो कि मेरे स्थानीय होस्ट LM Studio: https://lmstudio.ai पर चल रहे Gemma 4 12B को प्रॉम्प्ट करने का तरीका है:","LLM का Python API पहले आपसे एक सत्र बनाने और फिर एक समय में एक संदेश भेजने की मांग करता था। यह LLM की वास्तविक विशेषताओं का एक सार है, जिसमें प्रत्येक अनुरोध में पिछले सभी संदेशों का पूरा इतिहास शामिल होता है। कुछ अधिक उन्नत स्थितियों के लिए, यह सार अवरोध बनने लगता है, इसलिए नए संस्करण ने model.prompt(messages=[]) पैरामीटर पेश किया, जिसे इस तरह इस्तेमाल किया जा सकता है:","LLM पहले हर प्रॉम्प्ट से स्ट्रिंग का एक इटेरेबल सीक्वेंस वापस करता था। जब मॉडल स्ट्रिंग प्रतिक्रिया लौटाता था, यह अच्छा काम करता था, लेकिन यह पूर्वानुमान नहीं लगाया जा सकता था कि मॉडल किस अजीब रूप में विकसित होगा। आजकल, कई मॉडल तर्क पाठ, आउटपुट स्ट्रिंग, टूल कॉल, और यहां तक कि इमेज अटैचमेंट का मिश्रित कंटेंट लौटाते हैं। LLM 0.32 में, आप इस विधि का उपयोग कर सकते हैं: https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events","इन सुविधाओं को मिलाकर, हम अंततः एक मजबूत अर्ध-मानक OpenAI चैट पूर्णता API कार्यान्वयन प्रदान कर सकते हैं, जिसे मैंने llm-chat-completions-server के रूप में प्लगइन के रूप में जारी किया है: https://github.com/simonw/llm-chat-completions-server","अब, आप इस सर्वर के माध्यम से नए llm openai endpoint कमांड का उपयोग करके LLM पर प्रॉम्प्ट चला सकते हैं!","इस प्रकार के API की सबसे बड़ी चुनौती लॉगिंग में है। यदि हमें हर अनुरोध के लिए संदेश अनुक्रम जोड़ने के पैटर्न का समर्थन करना है, तो आदर्श रूप से हर सत्र के लिए सभी पुनरावर्ती JSON को रिकॉर्ड करने से बचा जा सकता है।","समाधान नया कंटेंट अड्रेस करने योग्य संदेश भंडारण है: https://llm.datasette.io/en/stable/logging.html#the-message-store, Git मॉडेलिंग का अनुकरण करता है। आप दस्तावेज़ में नया स्कीमा देख सकते हैं: https://llm.datasette.io/en/stable/logging.html#sql-schema, लेकिन llm logs और llm logs --json कमांड दोनों को अपग्रेड किया गया है, और यह फॉर्मेट आसानी से उपयोग करने योग्य रूप में वापस परिवर्तित किया जा सकता है।","इस रिलीज में और भी बहुत सारी सामग्री है। 0.32 रिलीज़ नोट्स: https://llm.datasette.io/en/stable/changelog.html#v0-32 काफी व्यापक हैं, 0.32rc2: https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30, 0.32rc: https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30, 0.32a3: https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09, 0.32a2: https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12 और 0.32a0: https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 के नोट्स किसी भी छूट को पूरा कर सकते हैं।","वर्तमान LLM प्लगइन्स को जारी रखा जा सकता है, लेकिन अतिरिक्त मॉडल प्रदान करने वाले प्लगइन्स को नए स्ट्रीम इवेंट सिस्टम में पूरी तरह से भाग लेने के लिए 0.32 में अपडेट करना होगा। दस्तावेज़ में संरचित संदेश और स्ट्रीमिंग इवेंट्स का उपयोग करके प्लगइन्स को लागू करने के बारे में मार्गदर्शन है: https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events।","मैंने अपने कुछ प्लगइन्स को अपडेट किया:","此次发布的许多底层工具更改是由 Datasette Agent：https://agent.datasette.io/ 的需求驱动的。当我开始开发 LLM 时，“代理”这个术语的定义非常模糊，我当时拒绝使用它。到 2025 年 9 月：https://simonwillison.net/2025/Sep/18/agents/ 我接受了这样一个观点：“LLM 代理通过循环运行工具以实现目标” 已经足够成熟，我可以完全停止回避这个术语。","टूलचेन अब मानव अनुमोदन की प्रतीक्षा करने के लिए रोक सकता है: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause, और संग्रहीत संदेश इतिहास से पुनः आरंभ कर सकता है: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume —— ये दोनों ही Datasette Agent के लिए आवश्यक हैं।","आज जब मैंने LLM देखा, तो यह मुझे एक एजेंट (agent) की तरह रूप दिखाने लगा। एक बहुत ही शानदार बात यह है कि आप एक कमांड लाइन टूल का उपयोग करके विभिन्न स्रोतों से, विभिन्न मॉडलों का उपयोग करके विभिन्न टूल्स को मिला सकते हैं, एक लाइन कोड में यह सब हो जाता है, और इसमें एक पर्याप्त मजबूत Python लाइब्रेरी भी शामिल है, जिससे Datasette Agent (https://agent.datasette.io/) और llm-coding-agent (https://github.com/simonw/llm-coding-agent) जैसी सिस्टम बनाई जा सकती हैं।","शायद अगली संस्करण की LLM 'प्रतिनिधि' की अवधारणा को सीधे मुख्य पुस्तकालय में शामिल कर देगी। मैं अभी भी यह समझने की कोशिश कर रहा हूँ कि वह कैसा होगा।","यह LLM का नया संस्करण है, जिसे Simon Willison ने 4 अगस्त 2026 को जारी किया: इसमें तर्क अनुवर्ती, OpenAI प्रतिक्रियाएँ, सर्वर-साइड टूल और अधिक बुद्धिमान लॉगिंग के लिए समर्थन जोड़ा गया है: /2026/Aug/4/।","श्रृंखला लेख \"बड़ी भाषा मॉडल (LLM) के नए संस्करण की रिलीज़\" का एक हिस्सा: /series/llm-releases/","पिछला लेख: स्टेटलेस MCP ने मेरी रुचि फिर से जगाई (और mcp-explorer और datasette-mcp को प्रेरित किया): /2026/Jul/31/stateless-mcp/","प्रति माह मुझे 10 डॉलर का समर्थन करें, और आप विशेष ईमेल सारांश प्राप्त करेंगे, जिसमें उस महीने के सबसे महत्वपूर्ण LLM समाचारों का संकलन होगा।"],"bodyTranslated":true,"bodySourceHash":"180cd773be407152","bodyTranslatedAt":"2026-08-08T12:25:08.313Z"},"it":{"title":"Rilasciato LLM 0.32: aggiunte traiettorie di inferenza, risposte OpenAI, strumenti lato server e logging più intelligente","summary":"Simon Willison ha rilasciato LLM 0.32, la versione nuova più importante dal lancio del progetto. La nuova versione supporta la visualizzazione delle traiettorie di inferenza, strumenti lato server, l'API OpenAI Responses e utilizza di default il modello Luna GPT-5.6.","category":"产品更新","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Rilasciato LLM 0.32: aggiunte traiettorie di inferenza, risposte OpenAI, strumenti lato server e logging più intelligente - Aioga Notizie IA","description":"Simon Willison ha rilasciato LLM 0.32, la versione nuova più importante dal lancio del progetto. La nuova versione supporta la visualizzazione delle traiettorie di inferenza, strum...","url":"https://www.aioga.com/it/news/cmsfdkkcs1ocxro2ed4z3s131/","contentTranslated":true,"sourceHash":"a2973b505f038783","translatedAt":"2026-08-05T01:23:20.054Z","articleBody":["Questa mattina ho rilasciato LLM 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32, che è la versione più importante di LLM dal primo rilascio del progetto. La nuova versione include il supporto per tracce di ragionamento visibili, strumenti per provider lato server, un registro SQLite a indirizzamento dei contenuti riprogettato, nuovi modelli e nuove funzionalità abilitate tramite l'API OpenAI Responses. Ho anche rilasciato una nuova versione del plugin llm-anthropic: https://github.com/simonw/llm-anthropic, anch'esso con molte novità.","Eseguire LLM su un modello di ragionamento ora mostrerà le loro tracce di ragionamento nell'output di errore standard, così puoi vedere il contenuto del loro \"pensiero\", mentre queste informazioni non saranno incluse nell'output standard che potresti inviare ad altri strumenti. Usare -R/--hide-reasoning può disattivare questa funzione.","LLM supporta nativamente la serie di modelli GPT-5.6, e il nuovo modello predefinito per i \"prompt\" di llm è ora il GPT-5.6 Luna, economico ma potente.","Le chiamate LLM possono ora utilizzare strumenti lato server da vari fornitori. OpenAI offre un ambiente di esecuzione del codice: https://llm.datasette.io/en/stable/openai-models.html#code-interpreter come strumento lato server; LLM può ora eseguire prompt che beneficiano di questo strumento, come mostrato di seguito:","OpenAI ha anche ottenuto uno strumento WebSearch: https://llm.datasette.io/en/stable/openai-models.html#web-search.","llm-anthropic: https://github.com/simonw/llm-anthropic Il plugin ha aggiunto WebSearch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search, WebFetch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch, CodeExecution: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution, e AnthropicMCP: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector, sembra essere così:","Questo causerà che Anthropic esegua chiamate MCP sul mio nuovo plugin datasette-mcp: https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp durante le interazioni di singola richiesta/risposta con la sua API.","Il nuovo comando llm openai endpoint fornisce uno strumento che consente di eseguire prompt su qualsiasi endpoint compatibile con OpenAI: https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it, tutto con un solo comando. Queste operazioni non vengono registrate, quindi è uno strumento comodo per eseguire prompt una tantum su qualsiasi oggetto che utilizza il linguaggio universale del mondo API LLM.","Ecco come lo utilizzo, attraverso uvx (senza necessità di installare LLM) e combinandolo con il plugin strumento llm-tools-quickjs: https://github.com/simonw/llm-tools-quickjs, per dare prompt a Gemma 4 12B che gira sul mio host locale LM Studio: https://lmstudio.ai:","L'API Python di LLM richiedeva in passato di creare una sessione e inviare un messaggio alla volta. Questa è un'astrazione delle vere caratteristiche di LLM, in cui ogni richiesta porta l'intero storico dei messaggi precedenti. Per alcuni casi più avanzati, questa astrazione cominciava a diventare un ostacolo, quindi le nuove versioni hanno introdotto il parametro model.prompt(messages=[]), che può essere utilizzato in questo modo:","Prima, gli LLM restituivano una sequenza iterabile di stringhe per ogni prompt. Questo funzionava bene quando il modello restituiva risposte sotto forma di stringa, ma non era possibile prevedere le strane forme che il modello poteva evolvere. Oggi, molti modelli restituiscono contenuti misti di testo deduttivo, stringhe di output, chiamate a strumenti e persino allegati di immagini. In LLM 0.32, puoi invece usare questo metodo: https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events","Combinando queste funzionalità, finalmente possiamo offrire un'implementazione robusta dell'API di completamento chat OpenAI semi-standard, che ho già pubblicato come plugin llm-chat-completions-server: https://github.com/simonw/llm-chat-completions-server","Ora, puoi utilizzare il nuovo comando endpoint llm openai tramite questo server per eseguire prompt su LLM!","La sfida maggiore di questo tipo di API riguarda la registrazione dei log. Se vogliamo supportare un modello in cui ogni richiesta aggiunge una sequenza di messaggi, idealmente si potrebbe evitare di registrare tutti i JSON duplicati per ogni sessione.","La soluzione è un nuovo archivio di messaggi indirizzabile per contenuti: https://llm.datasette.io/en/stable/logging.html#the-message-store, modellato sul Git. Puoi vedere il nuovo schema nella documentazione: https://llm.datasette.io/en/stable/logging.html#sql-schema, ma i comandi llm logs e llm logs --json sono stati entrambi aggiornati per convertire questo formato in una forma più facile da usare.","Questa versione contiene ancora più contenuti. Note di rilascio 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32 piuttosto complete, 0.32rc2: https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30, 0.32rc: https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30, 0.32a3: https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09, 0.32a2: https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12 e 0.32a0: https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 possono colmare eventuali omissioni.","I plugin LLM esistenti dovrebbero poter continuare a essere utilizzati, ma i plugin che offrono modelli aggiuntivi devono essere aggiornati alla versione 0.32 per partecipare pienamente al nuovo sistema di eventi in streaming. La documentazione contiene linee guida sull'uso di messaggi strutturati e eventi in streaming per implementare plugin: https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events.","Ho aggiornato alcuni dei miei plugin:","Molti dei cambiamenti agli strumenti di basso livello rilasciati questa volta sono stati guidati dalle esigenze di Datasette Agent: https://agent.datasette.io/. Quando ho iniziato a sviluppare LLM, il termine “agente” era molto vago, e all’epoca rifiutavo di usarlo. Entro settembre 2025: https://simonwillison.net/2025/Sep/18/agents/ ho accettato il punto di vista secondo cui “gli agenti LLM raggiungono obiettivi eseguendo strumenti in ciclo”, che era ormai abbastanza maturo da permettermi di smettere completamente di evitare questo termine.","La catena di strumenti può ora essere messa in pausa in attesa di approvazione manuale: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause, e può essere ripresa dalla cronologia dei messaggi memorizzata: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume — entrambe queste funzionalità sono necessarie per Datasette Agent.","Oggi guardando LLM, inizia a mostrarsi, a mio avviso, in una forma molto simile a un agente. C'è un aspetto molto interessante: puoi usare uno strumento da linea di comando per combinare strumenti provenienti da diverse fonti e che utilizzano modelli differenti, il tutto con una sola riga di codice, e include anche una libreria Python abbastanza potente per costruire sistemi come Datasette Agent (https://agent.datasette.io/) e llm-coding-agent (https://github.com/simonw/llm-coding-agent).","Forse la prossima versione del LLM integrerà direttamente il concetto di \"agente\" nella libreria principale. Sto ancora cercando di capire come sarà.","Questa è la nuova versione di LLM, rilasciata da Simon Willison il 4 agosto 2026: aggiunto il supporto per il tracciamento del ragionamento, le risposte di OpenAI, gli strumenti lato server e una registrazione dei log più intelligente: /2026/Aug/4/.","Parte della serie di articoli \"Nuove versioni dei grandi modelli linguistici (LLM)\": /series/llm-releases/","Articolo precedente: MCP senza stato ha riacceso il mio interesse (e ha ispirato mcp-explorer e datasette-mcp): /2026/Jul/31/stateless-mcp/","Sponsorizzami 10 dollari al mese e riceverai un riepilogo email selezionato, che riassume le novità più importanti sugli LLM del mese."],"bodyTranslated":true,"bodySourceHash":"180cd773be407152","bodyTranslatedAt":"2026-08-08T12:25:53.151Z"},"nl":{"title":"LLM 0.32 uitgebracht: toegevoegde inferentietrajecten, OpenAI-reacties, server-side tools en slimmer logging","summary":"Simon Willison bracht LLM 0.32 uit, de belangrijkste nieuwe versie sinds de lancering van het project. De nieuwe versie ondersteunt het weergeven van inferentietrajecten, server-side tools, de OpenAI Responses API en gebruikt standaard het GPT-5.6 Luna-model.","category":"产品更新","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"LLM 0.32 uitgebracht: toegevoegde inferentietrajecten, OpenAI-reacties, server-side tools en slimmer logging - Aioga AI-nieuws","description":"Simon Willison bracht LLM 0.32 uit, de belangrijkste nieuwe versie sinds de lancering van het project. De nieuwe versie ondersteunt het weergeven van inferentietrajecten, server-si...","url":"https://www.aioga.com/nl/news/cmsfdkkcs1ocxro2ed4z3s131/","contentTranslated":true,"sourceHash":"a2973b505f038783","translatedAt":"2026-08-05T01:23:19.881Z","articleBody":["Ik heb vanmorgen LLM 0.32 uitgebracht: https://llm.datasette.io/en/stable/changelog.html#v0-32, dit is de belangrijkste nieuwe versie van LLM sinds de eerste release van het project. De nieuwe versie bevat ondersteuning voor zichtbare redeneertrajecten, server-side provider tools, herontworpen content-adresserende SQLite-logboeken, nieuwe modellen en nieuwe functies ingeschakeld via de OpenAI Responses API. Ik heb ook een nieuwe versie van de llm-anthropic plugin uitgebracht: https://github.com/simonw/llm-anthropic, die zelf ook veel updates bevat.","Het uitvoeren van LLM op een redeneermodel zal nu hun redeneersporen weergeven op de standaardfoutuitvoer, zodat je hun 'gedachten' kunt zien, terwijl deze informatie niet wordt opgenomen in de standaarduitvoer die je mogelijk naar andere tools stuurt. Gebruik -R/--hide-reasoning om deze functie uit te schakelen.","LLM ondersteunt out-of-the-box de GPT-5.6 modelreeks, en het nieuwe standaardmodel voor llm \"prompt\" dat nu wordt gebruikt is de goedkope maar krachtige GPT-5.6 Luna.","LLM-aanroepen kunnen nu server-side tools van verschillende aanbieders gebruiken. OpenAI biedt een code-uitvoeringsomgeving aan: https://llm.datasette.io/en/stable/openai-models.html#code-interpreter als server-side tool; LLM kan nu prompts uitvoeren die van deze tool profiteren, zoals hieronder getoond:","OpenAI heeft ook een WebSearch gekregen: https://llm.datasette.io/en/stable/openai-models.html#web-search hulpmiddel.","llm-anthropic: https://github.com/simonw/llm-anthropic De plug-in heeft WebSearch toegevoegd: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search, WebFetch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch, CodeExecution: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution, en AnthropicMCP: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector, het ziet er als volgt uit:","Dit zal ertoe leiden dat Anthropic bij een enkele verzoek-/antwoordinteractie met zijn API een MCP-aanroep uitvoert op mijn nieuwe datasette-mcp: https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp plugin.","De nieuwe llm OpenAI-endpointopdracht biedt een hulpmiddel waarmee u prompts kunt uitvoeren op elk OpenAI-compatibel endpoint: https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it, met slechts één regel commando. Deze bewerkingen worden niet geregistreerd, dus het is een handig hulpmiddel om eenmalige prompts uit te voeren op elk object dat de LLM API wereldtaal gebruikt.","Dit is hoe ik het gebruik, via uvx (zonder LLM te installeren) en in combinatie met de llm-tools-quickjs: https://github.com/simonw/llm-tools-quickjs tool-plugin, om prompts te geven aan Gemma 4 12B die draait op mijn lokale host LM Studio: https://lmstudio.ai","De Python API van LLM vereiste vroeger dat je een sessie aanmaakte en vervolgens één bericht tegelijk verzond. Dit is een abstractie van de werkelijke kenmerken van LLM, waarbij elk verzoek de volledige geschiedenis van alle voorgaande berichten meedraagt. Voor sommige meer geavanceerde situaties begon deze abstractie een obstakel te worden, daarom introduceerde de nieuwe versie de parameter model.prompt(messages=[]) die als volgt kan worden gebruikt:","Vroeger zou LLM een iterabele reeks van strings teruggeven voor elke prompt. Dit werkte prima wanneer het model een stringrespons teruggaf, maar het was onmogelijk te voorspellen welke vreemde vormen het model zou kunnen aannemen. Tegenwoordig geven veel modellen een mix van redeneerteksten, uitvoerstrings, tool-aanroepen en zelfs afbeeldingsbijlagen terug. In LLM 0.32 kun je in plaats daarvan deze methode gebruiken: https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events","In combinatie met deze functies kunnen we eindelijk een robuuste semi-standaard OpenAI chat voltooiing API-implementatie aanbieden, die ik al heb uitgebracht als llm-chat-completions-server: https://github.com/simonw/llm-chat-completions-server plugin:","Nu kunt u via deze server het nieuwe llm openai endpoint-commando gebruiken om prompts op de LLM uit te voeren!","Het grotere probleem bij dit type API is het loggen. Als we het patroon willen ondersteunen waarbij bij elk verzoek een berichtreeks wordt toegevoegd, kan het ideaal zijn om te voorkomen dat alle herhaalde JSON voor elke sessie wordt geregistreerd.","De oplossing is een nieuw inhoud-adresserend berichtopslag: https://llm.datasette.io/en/stable/logging.html#the-message-store, gemodelleerd naar Git. Je kunt het nieuwe schema in de documentatie zien: https://llm.datasette.io/en/stable/logging.html#sql-schema, maar de opdrachten llm logs en llm logs --json zijn beide bijgewerkt, zodat dit formaat kan worden omgezet in een vorm die gemakkelijk te gebruiken is.","Deze release bevat nog meer inhoud. Release-opmerkingen voor 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32 zijn vrij uitgebreid, 0.32rc2: https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30, 0.32rc: https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30, 0.32a3: https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09, 0.32a2: https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12 en 0.32a0: https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 kunnen eventuele hiaten vullen.","Bestaande LLM-plugins zouden nog steeds bruikbaar moeten zijn, maar plugins die extra modellen aanbieden moeten worden geüpgraded naar 0.32 om volledig mee te kunnen doen aan het nieuwe streaming-event-systeem. In de documentatie is er een gids over het gebruik van gestructureerde berichten en streaming-events om plugins te implementeren: https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events.","Ik heb enkele van mijn eigen plug-ins bijgewerkt:","Veel van de onderliggende toolswijzigingen die in deze release zijn doorgevoerd, werden gedreven door de behoeften van Datasette Agent: https://agent.datasette.io/. Toen ik begon met het ontwikkelen van LLM, was de definitie van de term 'agent' erg vaag, en toen weigerde ik het te gebruiken. In september 2025: https://simonwillison.net/2025/Sep/18/agents/ accepteerde ik het standpunt dat 'LLM-agents hun tools herhaaldelijk gebruiken om doelen te bereiken' al voldoende volwassen was, en ik kon volledig stoppen met het vermijden van deze term.","De toolchain kan nu worden gepauzeerd om te wachten op handmatige goedkeuring: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause, en kan worden hervat vanuit de opgeslagen berichtgeschiedenis: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume — beide zijn vereist door Datasette Agent.","Vandaag keek ik naar LLM, en het begint voor mij een vorm aan te nemen die erg op een agent lijkt. Een heel mooi aspect is dat je met een commandoregeltool allerlei tools uit verschillende bronnen en met verschillende modellen kunt combineren, met slechts één regel code, en het omvat ook een krachtige Python-bibliotheek waarmee systemen zoals Datasette Agent (https://agent.datasette.io/) en llm-coding-agent (https://github.com/simonw/llm-coding-agent) gebouwd kunnen worden.","Misschien zal de volgende versie van LLM het concept van 'agent' rechtstreeks in de kernbibliotheek opnemen. Ik probeer nog steeds uit te zoeken hoe dat eruit zou zien.","Dit is de nieuwe versie van LLM, uitgebracht door Simon Willison op 4 augustus 2026: met toegevoegde ondersteuning voor redeneer-tracering, OpenAI-respons, server-side tools en slimmere logregistratie: /2026/Aug/4/.","Serie artikelen 'Nieuwe versies van grote taalmodellen (LLM) uitgebracht': /series/llm-releases/","Vorige artikel: Stateless MCP heeft opnieuw mijn interesse gewekt (en inspireerde mcp-explorer en datasette-mcp): /2026/Jul/31/stateless-mcp/","Door mij elke maand $10 te sponsoren, ontvangt u een zorgvuldig samengestelde e-mailsamenvatting met de belangrijkste LLM-ontwikkelingen van die maand."],"bodyTranslated":true,"bodySourceHash":"180cd773be407152","bodyTranslatedAt":"2026-08-08T12:25:51.884Z"},"tr":{"title":"LLM 0.32 yayınlandı: çıkarım yörüngeleri, OpenAI Yanıtları, sunucu tarafı araçlar ve daha akıllı kayıt eklendi","summary":"Simon Willison, projenin çıkışından bu yana en önemli yeni sürüm olan LLM 0.32'yi yayımladı. Yeni sürüm, çıkarım yörüngelerinin gösterilmesini, sunucu tarafı araçları, OpenAI Responses API'sini desteklemekte ve varsayılan olarak GPT-5.6 Luna modelini kullanmaktadır.","category":"产品更新","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"LLM 0.32 yayınlandı: çıkarım yörüngeleri, OpenAI Yanıtları, sunucu tarafı araçlar ve daha akıllı kayıt eklendi - Aioga AI Haberleri","description":"Simon Willison, projenin çıkışından bu yana en önemli yeni sürüm olan LLM 0.32'yi yayımladı. Yeni sürüm, çıkarım yörüngelerinin gösterilmesini, sunucu tarafı araçları, OpenAI Respo...","url":"https://www.aioga.com/tr/news/cmsfdkkcs1ocxro2ed4z3s131/","contentTranslated":true,"sourceHash":"a2973b505f038783","translatedAt":"2026-08-05T01:23:29.606Z","articleBody":["Bu sabah LLM 0.32'yi yayınladım: https://llm.datasette.io/en/stable/changelog.html#v0-32, bu, projenin ilk çıkışından bu yana LLM'nin en önemli yeni sürümü. Yeni sürüm, görünür çıkarım yörüngelerini desteklemek, sunucu tarafı sağlayıcı araçları, SQLite günlüklerini ele alan yeniden tasarlanmış içerik, yeni modeller ve OpenAI Yanıtlar API'si ile etkinleştirilen yeni özellikler içermektedir. Ayrıca llm-anthropic eklentisinin yeni bir sürümünü de yayınladım: https://github.com/simonw/llm-anthropic, kendisi de birçok güncellemeye sahip.","LLM'leri çıkarım modellerinde çalıştırmak artık onların çıkarım izlerini standart hata çıkışına gösterecek, böylece onların 'düşünme' içeriğini görebilirsiniz ve bu bilgiler diğer araçlara iletebileceğiniz standart çıktıya dahil edilmez. Bu özelliği kapatmak için -R/--hide-reasoning kullanabilirsiniz.","LLM, kutudan çıkar çıkmaz GPT-5.6 model serisini destekliyor ve şimdi llm 'prompt' için kullanılan yeni varsayılan model, ucuz ama güçlü GPT-5.6 Luna'dır.","LLM çağrıları artık çeşitli sağlayıcılardan gelen sunucu tarafı araçlarını kullanabilir. OpenAI, sunucu tarafı araç olarak bir kod yürütme ortamı sunar: https://llm.datasette.io/en/stable/openai-models.html#code-interpreter; LLM şimdi bu araçtan yararlanan istemleri çalıştırabilir, aşağıdaki gibi:","OpenAI ayrıca bir WebSearch aracı elde etti: https://llm.datasette.io/en/stable/openai-models.html#web-search.","llm-anthropic：https://github.com/simonw/llm-anthropic eklentisi WebSearch：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search, WebFetch：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch, CodeExecution：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution ve AnthropicMCP：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector kullanıcıya şöyle görünüyor:","Bu, Anthropic'in API ile yaptığı tek seferlik istek/yanıt etkileşimlerinde, benim yeni datasette-mcp: https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp eklentimde MCP çağrısı yapmasına neden olacak.","Yeni llm OpenAI uç noktası komutu, herhangi bir OpenAI uyumlu uç noktada istem çalıştırmak için bir araç sunuyor: https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it, tek bir komut satırıyla tamamlanabilir. Bu işlemler kaydedilmez, bu nedenle LLM API'sini kullanan herhangi bir evrensel dile sahip nesne üzerinde tek seferlik istemleri çalıştırmak için pratik bir araçtır.","Bu, bunu nasıl kullandığım: uvx (LLM yüklemeden) ve llm-tools-quickjs: https://github.com/simonw/llm-tools-quickjs araç eklentisini kullanarak, yerel bilgisayarımda çalışan LM Studio: https://lmstudio.ai'deki Gemma 4 12B üzerinde prompt verme şeklim:","LLM'nin Python API'si eskiden sizden bir oturum oluşturmanızı ve ardından mesajları tek tek göndermenizi gerektiriyordu. Bu, her isteğin önceki tüm mesajların tam geçmişini taşıdığı, LLM'nin gerçek özelliklerinin bir soyutlamasıdır. Bazı daha gelişmiş durumlarda, bu soyutlama bir engel haline gelmeye başlar, bu yüzden yeni sürüm model.prompt(messages=[]) parametresini tanıttı ve bunu şöyle kullanabilirsiniz:","LLM önce her istemden bir dizi yinelenebilir string döndürürdü. Model string yanıtlar verdiğinde bu çok iyi çalışıyordu, ancak modelin geliştirebileceği tuhaf biçimleri tahmin etmek mümkün değildi. Günümüzde birçok model akıl yürütme metni, çıktı stringi, araç çağrıları ve hatta resim ekleri gibi karışık içerikler döndürüyor. LLM 0.32'de bunun yerine şu yöntemi kullanabilirsin: https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events","Bu özellikleri birleştirerek, sonunda sağlam bir yarı standart OpenAI sohbet tamamlama API uygulaması sunabiliriz, bunu llm-chat-completions-server olarak yayınladım: https://github.com/simonw/llm-chat-completions-server eklenti olarak:","Artık bu sunucu üzerinden yeni llm openai uç noktası komutunu kullanarak LLM üzerinde istem çalıştırabilirsiniz!","Bu tür bir API'nin daha büyük zorluğu günlük kaydıdır. Eğer her istekte mesaj dizisini ekleyen bir modeli desteklemek istiyorsak, ideal olarak her oturum için tüm tekrar eden JSON'ları kaydetmekten kaçınabiliriz.","Çözüm, yeni içerik adreslenebilir mesaj depolamasıdır: https://llm.datasette.io/en/stable/logging.html#the-message-store, Git modellemesini taklit eder. Yeni şemayı belgelerde görebilirsiniz: https://llm.datasette.io/en/stable/logging.html#sql-schema, ancak llm logs ve llm logs --json komutları güncellendi ve bu formatı tekrar kullanımı kolay bir forma dönüştürebilir.","Bu sürümde daha fazla içerik var. 0.32 sürüm notları: https://llm.datasette.io/en/stable/changelog.html#v0-32 oldukça kapsamlı, 0.32rc2: https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30, 0.32rc: https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30, 0.32a3: https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09, 0.32a2: https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12 ve 0.32a0: https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 notları herhangi bir eksikliği tamamlayabilir.","Mevcut LLM eklentileri kullanılmaya devam edebilir, ancak ek modeller sunan eklentiler, yeni akış olayları sistemine tam olarak katılabilmek için 0.32 sürümüne yükseltilmelidir. Belgelerde, yapılandırılmış mesajlar ve akış olayları kullanarak eklentilerin nasıl uygulanacağına dair bir kılavuz bulunmaktadır: https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events.","Kendi eklentilerimden bazılarını güncelledim:","Bu sürümdeki birçok alt seviye araç değişikliği, Datasette Agent: https://agent.datasette.io/ taleplerinden kaynaklanmıştır. LLM geliştirmeye başladığımda, “ajan” teriminin tanımı oldukça belirsizdi ve o zamanlar kullanmaktan kaçındım. 2025 yılı Eylül ayına kadar: https://simonwillison.net/2025/Sep/18/agents/ böyle bir görüşü kabul ettim: “LLM ajanları, hedeflere ulaşmak için araçları döngüsel olarak çalıştırır” bu düşünce yeterince olgunlaşmıştı ve terimi tamamen kullanmaktan kaçınmayı bırakabilirdim.","Araç zinciri artık manuel onayı beklemek için duraklatılabilir: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause ve saklanan mesaj geçmişinden devam edilebilir: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume — her ikisi de Datasette Agent için gereklidir.","Bugün LLM'ye baktım, bana göre artık oldukça ajan (agent) benzeri bir form almaya başladı. Çok hoş bir yanı, farklı kaynaklardan gelen ve farklı modeller kullanan çeşitli araçları bir komut satırı aracıyla karıştırıp eşleştirebilmeniz, tek bir satır kodla halledebiliyorsunuz, ve ayrıca Datasette Agent (https://agent.datasette.io/) ve llm-coding-agent (https://github.com/simonw/llm-coding-agent) gibi sistemler kurmanıza olanak tanıyan yeterince güçlü bir Python kütüphanesi de içeriyor.","Belki bir sonraki LLM sürümü 'ajan' kavramını doğrudan çekirdek kütüphaneye entegre eder. Ben hâlâ bunun nasıl olacağını anlamaya çalışıyorum.","Bu, Simon Willison tarafından 4 Ağustos 2026'da yayınlanan LLM'in yeni sürümüdür: akıl yürütme takibi, OpenAI yanıtları, sunucu tarafı araçları ve daha akıllı günlük kaydı desteği eklendi: /2026/Aug/4/.","Büyük Dil Modelleri (LLM) Yeni Sürüm Yayınları serisinin bir bölümü: /series/llm-releases/","Önceki makale: Durumsuz MCP tekrar ilgimi çekti (ve mcp-explorer ile datasette-mcp'ye ilham verdi): /2026/Tem/31/stateless-mcp/","Her ay bana 10 dolar sponsor olursanız, o ayın en önemli LLM gelişmelerini derleyen seçilmiş e-posta özetini alabilirsiniz."],"bodyTranslated":true,"bodySourceHash":"180cd773be407152","bodyTranslatedAt":"2026-08-08T12:26:43.091Z"},"vi":{"title":"Phát hành LLM 0.32: bổ sung quỹ đạo suy luận, phản hồi OpenAI, công cụ phía máy chủ và ghi nhật ký thông minh hơn","summary":"Simon Willison đã phát hành LLM 0.32, phiên bản mới quan trọng nhất kể từ khi dự án ra mắt. Phiên bản mới hỗ trợ hiển thị quỹ đạo suy luận, công cụ phía máy chủ, API OpenAI Responses, và mặc định sử dụng mô hình GPT-5.6 Luna.","category":"产品更新","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Phát hành LLM 0.32: bổ sung quỹ đạo suy luận, phản hồi OpenAI, công cụ phía máy chủ và ghi nhật ký thông minh hơn - Tin tức AI Aioga","description":"Simon Willison đã phát hành LLM 0.32, phiên bản mới quan trọng nhất kể từ khi dự án ra mắt. Phiên bản mới hỗ trợ hiển thị quỹ đạo suy luận, công cụ phía máy chủ, API OpenAI Respons...","url":"https://www.aioga.com/vi/news/cmsfdkkcs1ocxro2ed4z3s131/","contentTranslated":true,"sourceHash":"a2973b505f038783","translatedAt":"2026-08-05T01:23:28.761Z","articleBody":["Sáng nay tôi đã phát hành LLM 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32, đây là phiên bản mới quan trọng nhất của LLM kể từ khi dự án được phát hành lần đầu. Phiên bản mới bao gồm hỗ trợ theo dõi suy luận có thể nhìn thấy, công cụ nhà cung cấp phía máy chủ, nhật ký SQLite theo địa chỉ nội dung được thiết kế lại, mô hình mới, và các tính năng mới được kích hoạt thông qua OpenAI Responses API. Tôi cũng đã phát hành phiên bản mới của plugin llm-anthropic: https://github.com/simonw/llm-anthropic, cũng có nhiều cập nhật lớn.","Chạy LLM trên mô hình suy luận bây giờ sẽ hiển thị các lộ trình suy luận của chúng vào đầu ra lỗi chuẩn, vì vậy bạn có thể thấy nội dung \"suy nghĩ\" của chúng, và thông tin này sẽ không được bao gồm trong đầu ra chuẩn mà bạn có thể gửi cho các công cụ khác. Sử dụng -R/--hide-reasoning có thể tắt tính năng này.","LLM hỗ trợ ngay lập tức dòng mô hình GPT-5.6, và mô hình mặc định mới khi sử dụng câu lệnh 'prompt' của llm hiện nay là GPT-5.6 Luna, giá rẻ nhưng mạnh mẽ.","Gọi LLM hiện có thể sử dụng các công cụ phía máy chủ từ nhiều nhà cung cấp khác nhau. OpenAI cung cấp một môi trường thực thi mã: https://llm.datasette.io/en/stable/openai-models.html#code-interpreter như một công cụ phía máy chủ; LLM hiện có thể chạy các prompt được hưởng lợi từ công cụ này, như sau:","OpenAI cũng đã có được một công cụ WebSearch: https://llm.datasette.io/en/stable/openai-models.html#web-search.","llm-anthropic: https://github.com/simonw/llm-anthropic Plugin đã thêm WebSearch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search, WebFetch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch, CodeExecution: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution, cũng như AnthropicMCP: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector, trông có vẻ như sau:","Điều này sẽ dẫn đến việc Anthropic thực hiện các cuộc gọi MCP đối với plugin datasette-mcp mới của tôi: https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp trong các tương tác yêu cầu/phản hồi một lần với API của họ.","Lệnh endpoint llm mới của OpenAI cung cấp một công cụ, có thể thực hiện các prompt đối với bất kỳ endpoint nào tương thích với OpenAI: https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it, chỉ với một dòng lệnh là xong. Những thao tác này sẽ không được ghi lại, do đó đây là một công cụ tiện lợi, có thể chạy prompt một lần đối với bất kỳ đối tượng nào sử dụng ngôn ngữ phổ quát của LLM API.","Đây là cách tôi sử dụng nó, thông qua uvx (không cần cài đặt LLM) và kết hợp với plugin công cụ llm-tools-quickjs: https://github.com/simonw/llm-tools-quickjs, để hướng dẫn Gemma 4 12B chạy trên máy chủ cục bộ LM Studio của tôi: https://lmstudio.ai","API Python của LLM trước đây yêu cầu bạn tạo một phiên, sau đó gửi một tin nhắn một lần. Đây là một trừu tượng hóa thực sự của các đặc tính của LLM, mỗi yêu cầu đều mang theo toàn bộ lịch sử của tất cả các tin nhắn trước đó. Đối với một số trường hợp nâng cao hơn, trừu tượng hóa này bắt đầu trở thành trở ngại, do đó phiên bản mới đã giới thiệu tham số model.prompt(messages=[]) có thể được sử dụng như sau:","Trước đây, LLM sẽ trả về một chuỗi các chuỗi có thể lặp lại từ mỗi lời nhắc. Khi mô hình trả về phản hồi dạng chuỗi, điều này hoạt động tốt, nhưng không thể dự đoán được những dạng kỳ lạ mà mô hình có thể phát triển. Ngày nay, nhiều mô hình trả về nội dung hỗn hợp bao gồm văn bản suy luận, chuỗi đầu ra, lời gọi công cụ, thậm chí là tệp đính kèm hình ảnh. Trong LLM 0.32, bạn có thể sử dụng phương pháp này: https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events","Kết hợp những chức năng này, chúng tôi cuối cùng có thể cung cấp một triển khai API hoàn tất trò chuyện OpenAI bán chuẩn ổn định, tôi đã phát hành nó dưới dạng plugin llm-chat-completions-server: https://github.com/simonw/llm-chat-completions-server","Bây giờ, bạn có thể sử dụng lệnh điểm cuối llm openai mới thông qua máy chủ này để chạy các lời nhắc trên LLM!","Thách thức lớn hơn của loại API này là ghi nhật ký. Nếu chúng ta muốn hỗ trợ mô hình thêm dãy thông điệp cho mỗi yêu cầu, lý tưởng là có thể tránh ghi lại toàn bộ JSON trùng lặp cho mỗi phiên.","Giải pháp là lưu trữ thông điệp có thể định vị nội dung mới: https://llm.datasette.io/en/stable/logging.html#the-message-store, mô phỏng mô hình Git. Bạn có thể xem mẫu mới trong tài liệu: https://llm.datasette.io/en/stable/logging.html#sql-schema, nhưng các lệnh llm logs và llm logs --json đều đã được nâng cấp, có thể chuyển đổi định dạng này trở lại dạng dễ sử dụng.","Bản phát hành này còn có nhiều nội dung hơn. Ghi chú phát hành 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32 khá đầy đủ, 0.32rc2: https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30, 0.32rc: https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30, 0.32a3: https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09, 0.32a2: https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12 và 0.32a0: https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 có thể bổ sung bất kỳ thiếu sót nào.","Các plugin LLM hiện có nên vẫn có thể tiếp tục sử dụng, nhưng các plugin cung cấp mô hình bổ sung cần nâng cấp lên phiên bản 0.32 để có thể tham gia đầy đủ vào hệ thống sự kiện luồng mới. Tài liệu có hướng dẫn về cách sử dụng tin nhắn có cấu trúc và sự kiện luồng để triển khai plugin: https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events.","Tôi đã cập nhật một số plugin của riêng tôi:","Nhiều thay đổi về công cụ cơ sở được phát hành lần này được thúc đẩy bởi nhu cầu từ Datasette Agent: https://agent.datasette.io/. Khi tôi bắt đầu phát triển LLM, thuật ngữ \"đại lý\" còn khá mơ hồ, và tôi lúc đó đã từ chối sử dụng nó. Đến tháng 9 năm 2025: https://simonwillison.net/2025/Sep/18/agents/ tôi đã chấp nhận quan điểm rằng \"Đại lý LLM đạt được mục tiêu thông qua việc chạy công cụ theo vòng lặp\" đã đủ trưởng thành, và tôi có thể hoàn toàn ngừng né tránh thuật ngữ này.","Chuỗi công cụ hiện có thể tạm dừng để chờ phê duyệt thủ công: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause, và khôi phục từ lịch sử tin nhắn đã lưu: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume —— cả hai đều là những gì Datasette Agent cần.","Hôm nay xem LLM, nó bắt đầu hiện ra trong mắt tôi với hình thức rất giống đại lý (agent). Có một điểm rất hay là bạn có thể dùng một công cụ dòng lệnh để kết hợp các công cụ từ các nguồn khác nhau, sử dụng các mô hình khác nhau, chỉ bằng một dòng lệnh là xong, và nó bao gồm một thư viện Python đủ mạnh để xây dựng các hệ thống như Datasette Agent (https://agent.datasette.io/) và llm-coding-agent (https://github.com/simonw/llm-coding-agent).","Có lẽ phiên bản LLM tiếp theo sẽ tích hợp trực tiếp khái niệm 'tác nhân' vào thư viện cốt lõi. Tôi vẫn đang cố gắng tìm hiểu xem điều đó sẽ trông như thế nào.","Đây là phiên bản mới của LLM, được Simon Willison phát hành vào ngày 4 tháng 8 năm 2026: bổ sung hỗ trợ theo dõi lý luận, phản hồi OpenAI, công cụ phía máy chủ và ghi nhật ký thông minh hơn: /2026/Aug/4/.","Một phần của loạt bài viết “Phát hành phiên bản mới của Mô hình Ngôn ngữ Lớn (LLM)”: /series/llm-releases/","Bài trước: MCP không trạng thái đã khiến tôi quan tâm trở lại (và truyền cảm hứng cho mcp-explorer và datasette-mcp): /2026/Jul/31/stateless-mcp/","Hàng tháng tài trợ cho tôi 10 đô la, bạn sẽ nhận được bản tóm tắt email được chọn lọc, tổng hợp những động thái quan trọng nhất của LLM trong tháng."],"bodyTranslated":true,"bodySourceHash":"180cd773be407152","bodyTranslatedAt":"2026-08-08T12:26:46.879Z"},"id":{"title":"LLM 0.32 dirilis: menambahkan lintasan inferensi, Respons OpenAI, alat sisi server, dan pencatatan yang lebih cerdas","summary":"Simon Willison merilis LLM 0.32, versi baru terpenting sejak peluncuran proyek tersebut. Versi baru mendukung tampilan lintasan inferensi, alat sisi server, OpenAI Responses API, dan secara default menggunakan model Luna GPT-5.6.","category":"产品更新","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"LLM 0.32 dirilis: menambahkan lintasan inferensi, Respons OpenAI, alat sisi server, dan pencatatan yang lebih cerdas - Berita AI Aioga","description":"Simon Willison merilis LLM 0.32, versi baru terpenting sejak peluncuran proyek tersebut. Versi baru mendukung tampilan lintasan inferensi, alat sisi server, OpenAI Responses API, d...","url":"https://www.aioga.com/id/news/cmsfdkkcs1ocxro2ed4z3s131/","contentTranslated":true,"sourceHash":"a2973b505f038783","translatedAt":"2026-08-05T01:23:37.444Z","articleBody":["Pagi ini saya merilis LLM 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32, ini adalah versi baru terpenting dari LLM sejak proyek ini pertama kali dirilis. Versi baru ini mencakup dukungan untuk jejak penalaran yang terlihat, alat penyedia sisi server, log SQLite yang diarahkan ulang dengan cara baru, model baru, serta fitur baru yang diaktifkan melalui OpenAI Responses API. Saya juga merilis versi baru plugin llm-anthropic: https://github.com/simonw/llm-anthropic, yang juga memiliki banyak pembaruan.","Menjalankan LLM pada model penalaran sekarang akan menampilkan jejak penalaran mereka ke output kesalahan standar, sehingga Anda bisa melihat isi 'pemikiran' mereka, dan informasi ini tidak akan termasuk dalam output standar yang mungkin Anda kirimkan ke alat lain. Menggunakan -R/--hide-reasoning dapat menonaktifkan fitur ini.","LLM langsung mendukung seri model GPT-5.6, dan model default baru untuk menggunakan \"prompt\" llm sekarang adalah GPT-5.6 Luna yang murah tetapi kuat.","Pemanggilan LLM sekarang dapat menggunakan alat sisi server dari berbagai penyedia. OpenAI menyediakan lingkungan eksekusi kode: https://llm.datasette.io/en/stable/openai-models.html#code-interpreter sebagai alat sisi server; LLM sekarang dapat menjalankan prompt yang memanfaatkan alat ini, seperti yang ditunjukkan di bawah ini:","OpenAI juga mendapatkan sebuah alat WebSearch: https://llm.datasette.io/en/stable/openai-models.html#web-search.","llm-anthropic: https://github.com/simonw/llm-anthropic Plugin ini menambahkan WebSearch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search, WebFetch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch, CodeExecution: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution, dan AnthropicMCP: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector, terlihat seperti ini:","Ini akan menyebabkan Anthropic melakukan panggilan MCP pada plugin datasette-mcp baru saya: https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp dalam interaksi permintaan/respons tunggal dengan API mereka.","Perintah endpoint llm OpenAI baru menyediakan sebuah alat yang dapat menjalankan prompt pada endpoint yang kompatibel dengan OpenAI: https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it, hanya dengan satu baris perintah. Operasi ini tidak dicatat, sehingga ini adalah alat yang praktis untuk menjalankan prompt satu kali pada objek mana pun yang menggunakan bahasa universal dunia LLM API.","Ini adalah cara saya menggunakannya, melalui uvx (tanpa perlu menginstal LLM) dan menggabungkannya dengan plugin alat llm-tools-quickjs: https://github.com/simonw/llm-tools-quickjs, untuk memberi prompt pada Gemma 4 12B yang berjalan di LM Studio saya sendiri: https://lmstudio.ai","API Python LLM sebelumnya mengharuskan Anda membuat sesi, kemudian mengirim satu pesan sekaligus. Ini adalah abstraksi dari fitur nyata LLM, di mana setiap permintaan membawa riwayat lengkap semua pesan sebelumnya. Untuk beberapa kasus yang lebih canggih, abstraksi ini mulai menjadi hambatan, sehingga versi baru memperkenalkan parameter model.prompt(messages=[]), yang dapat digunakan seperti ini:","Sebelumnya, LLM akan mengembalikan sebuah urutan iterable dari string untuk setiap prompt. Ini bekerja dengan baik saat model mengembalikan respons string, tetapi tidak dapat memprediksi bentuk aneh yang mungkin berkembang dari model. Saat ini, banyak model mengembalikan kombinasi teks penalaran, string output, panggilan alat, bahkan lampiran gambar. Di LLM 0.32, kamu bisa menggunakan metode ini sebagai gantinya: https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events","Dengan menggabungkan fitur-fitur ini, akhirnya kami dapat menyediakan implementasi API penyelesaian obrolan OpenAI setengah standar yang tangguh, yang telah saya rilis sebagai plugin llm-chat-completions-server: https://github.com/simonw/llm-chat-completions-server","Sekarang, Anda dapat menggunakan perintah endpoint llm openai baru melalui server ini untuk menjalankan prompt pada LLM!","Tantangan yang lebih besar dari jenis API ini terletak pada pencatatan log. Jika kita ingin mendukung pola penambahan urutan pesan pada setiap permintaan, secara ideal kita dapat menghindari pencatatan semua JSON yang sama untuk setiap sesi.","Solusinya adalah penyimpanan pesan yang dapat diakses berdasarkan konten baru: https://llm.datasette.io/en/stable/logging.html#the-message-store, meniru pemodelan Git. Kamu dapat melihat skema baru dalam dokumentasi: https://llm.datasette.io/en/stable/logging.html#sql-schema, tetapi perintah llm logs dan llm logs --json telah diperbarui, sehingga format tersebut dapat dikonversi kembali ke bentuk yang mudah digunakan.","Rilis ini juga memiliki lebih banyak konten. Catatan rilis 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32 cukup lengkap, 0.32rc2: https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30, 0.32rc: https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30, 0.32a3: https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09, 0.32a2: https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12 dan 0.32a0: https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 dapat mengisi setiap kekurangan.","Plugin LLM yang ada seharusnya masih bisa terus digunakan, tetapi plugin yang menyediakan model tambahan perlu diperbarui ke versi 0.32 agar bisa sepenuhnya berpartisipasi dalam sistem event streaming baru. Dokumentasi memiliki panduan tentang penggunaan pesan terstruktur dan event streaming untuk mengimplementasikan plugin: https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events.","Saya telah memperbarui beberapa plugin saya sendiri:","Banyak perubahan alat dasar yang dirilis kali ini didorong oleh kebutuhan Datasette Agent: https://agent.datasette.io/. Ketika saya mulai mengembangkan LLM, istilah 'agen' sangat ambigu, sehingga saat itu saya menolak menggunakannya. Pada September 2025: https://simonwillison.net/2025/Sep/18/agents/ saya menerima pandangan bahwa 'agen LLM menjalankan alat secara berulang untuk mencapai tujuan' sudah cukup matang, sehingga saya dapat sepenuhnya berhenti menghindari istilah ini.","Rantai alat sekarang dapat dijeda untuk menunggu persetujuan manual: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause, dan dipulihkan dari riwayat pesan yang disimpan: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume——kedua hal ini diperlukan oleh Datasette Agent.","Hari ini melihat LLM, bagi saya mulai menunjukkan bentuk yang sangat mirip dengan agen (agent). Ada satu hal yang menarik, yaitu Anda dapat menggunakan alat baris perintah untuk menggabungkan berbagai alat dari sumber berbeda dan menggunakan model yang berbeda, hanya dengan satu baris kode, dan itu termasuk pustaka Python yang cukup kuat untuk membangun sistem seperti Datasette Agent (https://agent.datasette.io/) dan llm-coding-agent (https://github.com/simonw/llm-coding-agent).","Mungkin versi LLM berikutnya akan langsung mengintegrasikan konsep 'agen' ke dalam pustaka inti. Saya masih mencoba mencari tahu seperti apa itu nantinya.","Ini adalah versi baru LLM, dirilis oleh Simon Willison pada 4 Agustus 2026: menambahkan dukungan untuk pelacakan penalaran, respons OpenAI, alat sisi server, serta pencatatan log yang lebih cerdas: /2026/Aug/4/.","Bagian dari seri artikel 'Rilis Versi Baru Model Bahasa Besar (LLM)': /series/llm-releases/","Artikel sebelumnya: MCP tanpa status kembali menarik minat saya (dan menginspirasi mcp-explorer dan datasette-mcp): /2026/Jul/31/stateless-mcp/","Mendukung saya $10 per bulan, Anda akan menerima ringkasan email pilihan yang merangkum perkembangan LLM terpenting bulan ini."],"bodyTranslated":true,"bodySourceHash":"180cd773be407152","bodyTranslatedAt":"2026-08-08T12:27:26.472Z"},"th":{"title":"LLM 0.32 เปิดตัว: เพิ่มเส้นทางการอนุมาน, OpenAI Responses, เครื่องมือฝั่งเซิร์ฟเวอร์ และการบันทึกที่ชาญฉลาดขึ้น","summary":"Simon Willison ได้ปล่อย LLM 0.32 ซึ่งเป็นเวอร์ชันใหม่ที่สําคัญที่สุดนับตั้งแต่เปิดตัวโครงการ เวอร์ชันใหม่รองรับการแสดงเส้นทางการอนุมาน เครื่องมือฝั่งเซิร์ฟเวอร์ API OpenAI Responses และใช้โมเดล GPT-5.6 Luna เป็นค่าเริ่มต้น","category":"产品更新","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"LLM 0.32 เปิดตัว: เพิ่มเส้นทางการอนุมาน, OpenAI Responses, เครื่องมือฝั่งเซิร์ฟเวอร์ และการบันทึกที่ชาญฉลาดขึ้น - ข่าว AI Aioga","description":"Simon Willison ได้ปล่อย LLM 0.32 ซึ่งเป็นเวอร์ชันใหม่ที่สําคัญที่สุดนับตั้งแต่เปิดตัวโครงการ เวอร์ชันใหม่รองรับการแสดงเส้นทางการอนุมาน เครื่องมือฝั่งเซิร์ฟเวอร์ API OpenAI Response...","url":"https://www.aioga.com/th/news/cmsfdkkcs1ocxro2ed4z3s131/","contentTranslated":true,"sourceHash":"a2973b505f038783","translatedAt":"2026-08-05T01:23:38.413Z","articleBody":["เช้านี้ฉันได้ปล่อย LLM 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32 ซึ่งเป็นเวอร์ชันใหม่ที่สำคัญที่สุดของ LLM นับตั้งแต่โครงการเปิดตัวครั้งแรก เวอร์ชันใหม่รวมถึงการรองรับเส้นทางการคิดที่มองเห็นได้ เครื่องมือผู้ให้บริการฝั่งเซิร์ฟเวอร์ การออกแบบใหม่ของบันทึก SQLite แบบเข้าถึงตามเนื้อหา โมเดลใหม่ และฟีเจอร์ใหม่ที่เปิดใช้งานผ่าน OpenAI Responses API ฉันยังได้ปล่อยเวอร์ชันใหม่ของปลั๊กอิน llm-anthropic: https://github.com/simonw/llm-anthropic โดยตัวมันเองก็มีการปรับปรุงมากมาย","การรัน LLM บนโมเดลการให้เหตุผลตอนนี้จะแสดงเส้นทางการให้เหตุผลของพวกมันไปยังเอาต์พุตข้อผิดพลาดมาตรฐาน ดังนั้นคุณสามารถเห็นเนื้อหาการ “คิด” ของพวกมัน ข้อมูลเหล่านี้จะไม่ถูกใส่ในเอาต์พุตมาตรฐานที่คุณอาจส่งไปยังเครื่องมืออื่น ๆ การใช้ -R/--hide-reasoning สามารถปิดฟังก์ชันนี้ได้","LLM เปิดใช้งานได้ทันทีรองรับชุดโมเดล GPT-5.6 และโมเดลใหม่เริ่มต้นสำหรับการใช้ \"prompt\" ของ llm ตอนนี้คือ GPT-5.6 Luna ราคาประหยัดแต่ทรงพลัง","การเรียกใช้งาน LLM ตอนนี้สามารถใช้เครื่องมือฝั่งเซิร์ฟเวอร์จากผู้ให้บริการต่าง ๆ OpenAI ให้สภาพแวดล้อมการประมวลผลโค้ด: https://llm.datasette.io/en/stable/openai-models.html#code-interpreter เป็นเครื่องมือฝั่งเซิร์ฟเวอร์; ตอนนี้ LLM สามารถเรียกใช้งานพรอมต์ที่ได้รับประโยชน์จากเครื่องมือนี้ได้ดังนี้:","OpenAI ยังได้รับเครื่องมือ WebSearch: https://llm.datasette.io/en/stable/openai-models.html#web-search","llm-anthropic: https://github.com/simonw/llm-anthropic ปลั๊กอินเพิ่ม WebSearch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search, WebFetch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch, CodeExecution: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution และ AnthropicMCP: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector ดูเหมือนจะเป็นแบบนี้:","สิ่งนี้จะทำให้ Anthropic ในการโต้ตอบแบบคำขอ/การตอบสนองเดียวกับ API ของมัน เรียกใช้ MCP บนปลั๊กอิน datasette-mcp ใหม่ของฉัน: https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp","คำสั่ง llm openai endpoint ใหม่มีเครื่องมือที่สามารถเรียกใช้คำสั่งกับเอ็นด์พอยต์ที่เข้ากันได้กับ OpenAI ได้: https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it ทำได้เพียงบรรทัดคำสั่งเดียว การดำเนินการเหล่านี้จะไม่ถูกบันทึก ดังนั้นจึงเป็นเครื่องมือที่สะดวกในการเรียกใช้คำสั่งครั้งเดียวสำหรับวัตถุใด ๆ ที่ใช้ภาษาในโลกสากลของ LLM API","นี่คือวิธีที่ฉันใช้มัน ผ่าน uvx (ไม่จำเป็นต้องติดตั้ง LLM) และรวมกับปลั๊กอินเครื่องมือ llm-tools-quickjs: https://github.com/simonw/llm-tools-quickjs ในการให้คำสั่งกับ Gemma 4 12B ที่ทำงานบนโฮสต์เครื่องของฉัน LM Studio: https://lmstudio.ai","Python API ของ LLM ก่อนหน้านี้ต้องให้คุณสร้างเซสชัน จากนั้นส่งข้อความทีละข้อความ นี่คือการสรุปลักษณะจริงของ LLM โดยแต่ละคำขอจะมีประวัติข้อความก่อนหน้าทั้งหมด สำหรับบางกรณีที่ซับซ้อนขึ้น การสรุปแบบนี้เริ่มกลายเป็นอุปสรรค ดังนั้นเวอร์ชันใหม่จึงแนะนำพารามิเตอร์ model.prompt(messages=[]) ซึ่งสามารถใช้ได้ดังนี้:","LLM ก่อนหน้านี้จะคืนลำดับที่สามารถวนซ้ำได้ของสตริงจากแต่ละพรอมต์ เมื่อโมเดลคืนการตอบสนองเป็นสตริง สิ่งนี้ใช้ได้ดี แต่ไม่สามารถทำนายรูปแบบแปลก ๆ ที่โมเดลอาจพัฒนาได้ทุกวันนี้ โมเดลหลายตัวคืนข้อความการใช้เหตุผล, สตริงเอาต์พุต, การเรียกเครื่องมือ หรือแม้แต่เนื้อหาที่ผสมกับไฟล์แนบภาพ ใน LLM 0.32 คุณสามารถใช้วิธีนี้แทน: https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events","รวมฟังก์ชันเหล่านี้แล้ว เราสามารถเสนอการใช้งาน API การสนทนาแบบ OpenAI ครึ่งมาตรฐานที่มั่นคงได้ในที่สุด ผมได้เผยแพร่เป็นปลั๊กอิน llm-chat-completions-server ที่นี่: https://github.com/simonw/llm-chat-completions-server","ตอนนี้ คุณสามารถใช้คำสั่ง llm openai endpoint ใหม่ผ่านเซิร์ฟเวอร์นี้เพื่อรันพรอมต์กับ LLM ได้แล้ว!","ความท้าทายที่ใหญ่กว่าใน API ประเภทนี้อยู่ที่การบันทึกล็อก หากเราต้องการสนับสนุนรูปแบบที่เพิ่มลำดับข้อความในทุกคำขอ อุดมคติคือต้องหลีกเลี่ยงการบันทึก JSON ซ้ำทั้งหมดสำหรับแต่ละเซสชัน","วิธีแก้ปัญหาคือการจัดเก็บข้อความที่สามารถระบุได้ด้วยเนื้อหาใหม่: https://llm.datasette.io/en/stable/logging.html#the-message-store โดยเลียนแบบการสร้างแบบจำลองของ Git คุณสามารถดูรูปแบบใหม่ในเอกสารได้ที่นี่: https://llm.datasette.io/en/stable/logging.html#sql-schema แต่คำสั่ง llm logs และ llm logs --json ได้รับการอัปเกรดแล้ว สามารถแปลงรูปแบบนี้กลับมาให้ง่ายต่อการใช้งานได้","การเปิดตัวครั้งนี้ยังมีเนื้อหาเพิ่มเติม รายละเอียดการออกเวอร์ชัน 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32 ครอบคลุมค่อนข้างครบถ้วน, 0.32rc2: https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30, 0.32rc: https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30, 0.32a3: https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09, 0.32a2: https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12 และ 0.32a0: https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 สามารถเติมเต็มสิ่งที่อาจตกหล่นได้","ปลั๊กอิน LLM ที่มีอยู่ในปัจจุบันควรใช้งานต่อไปได้ แต่ปลั๊กอินที่ให้โมเดลเพิ่มเติมจำเป็นต้องอัปเกรดเป็นเวอร์ชัน 0.32 จึงจะสามารถเข้าร่วมระบบเหตุการณ์สตรีมใหม่ได้อย่างสมบูรณ์ ในเอกสารมีคู่มือเกี่ยวกับการใช้ข้อความแบบมีโครงสร้างและเหตุการณ์สตรีมเพื่อสร้างปลั๊กอิน: https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events.","ฉันอัปเดตปลั๊กอินบางส่วนของตัวเอง:","การเปลี่ยนแปลงเครื่องมือระดับพื้นฐานหลายอย่างที่เปิดตัวในครั้งนี้ได้รับแรงขับเคลื่อนจากความต้องการของ Datasette Agent: https://agent.datasette.io/ เมื่อฉันเริ่มพัฒนา LLM คำว่า “ตัวแทน” ยังมีความหมายไม่ชัดเจน และฉันในตอนนั้นปฏิเสธที่จะใช้มัน จนถึงเดือนกันยายน 2025: https://simonwillison.net/2025/Sep/18/agents/ ฉันยอมรับมุมมองว่า “ตัวแทน LLM ทำงานโดยใช้เครื่องมือวนซ้ำเพื่อบรรลุเป้าหมาย” ซึ่งมีความชัดเจนเพียงพอที่ทำให้ฉันสามารถหยุดเลี่ยงการใช้คำนี้ได้โดยสิ้นเชิง","ตอนนี้เครื่องมือสามารถหยุดชั่วคราวเพื่อรอการอนุมัติจากมนุษย์ได้: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause และสามารถกู้คืนจากประวัติข้อความที่จัดเก็บไว้ได้: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume—ทั้งสองสิ่งนี้เป็นสิ่งที่ Datasette Agent จำเป็นต้องใช้.","วันนี้ดู LLM มันเริ่มแสดงให้เห็นรูปร่างที่คล้ายกับเอเจนต์ (agent) มากสำหรับฉัน สิ่งที่น่าสนใจอย่างหนึ่งคือ คุณสามารถใช้เครื่องมือบรรทัดคำสั่งในการรวมเครื่องมือต่าง ๆ ที่มาจากแหล่งที่มาและใช้โมเดลต่างกันเข้าด้วยกันได้ เพียงบรรทัดคำสั่งเดียวก็ทำได้ และมันรวมถึงไลบรารี Python ที่มีประสิทธิภาพเพียงพอที่จะสร้างระบบเช่น Datasette Agent (https://agent.datasette.io/) และ llm-coding-agent (https://github.com/simonw/llm-coding-agent) ได้","บางทีเวอร์ชันต่อไปของ LLM อาจรวมแนวคิดเรื่อง 'ตัวแทน' เข้ากับไลบรารีหลักโดยตรง ฉันยังคงพยายามทำความเข้าใจว่ามันจะเป็นอย่างไร","นี่คือเวอร์ชันใหม่ของ LLM โดย Simon Willison เผยแพร่เมื่อวันที่ 4 สิงหาคม 2026: เพิ่มการสนับสนุนการติดตามเหตุผล การตอบสนองจาก OpenAI เครื่องมือฝั่งเซิร์ฟเวอร์ และการบันทึกล็อกที่ชาญฉลาดมากขึ้น: /2026/Aug/4/.","ส่วนหนึ่งของบทความชุด “การเปิดตัวเวอร์ชันใหม่ของโมเดลภาษาขนาดใหญ่ (LLM)” : /series/llm-releases/","บทความก่อนหน้า: MCP แบบไม่มีสถานะ ทำให้ความสนใจของฉันกลับมา (และเป็นแรงบันดาลใจให้กับ mcp-explorer และ datasette-mcp): /2026/Jul/31/stateless-mcp/","สนับสนุนฉันเดือนละ 10 ดอลลาร์ แล้วคุณจะได้รับสรุปอีเมลที่คัดสรรมา ซึ่งรวมข้อมูลอัปเดต LLM ที่สำคัญที่สุดของเดือนนั้น"],"bodyTranslated":true,"bodySourceHash":"180cd773be407152","bodyTranslatedAt":"2026-08-08T12:27:36.195Z"},"pl":{"title":"Wydany LLM 0.32: dodane trajektorie wnioskowania, OpenAI Responses, narzędzia po stronie serwera oraz inteligentniejsze logowanie","summary":"Simon Willison wydał LLM 0.32, najważniejszą nową wersję od czasu uruchomienia projektu. Nowa wersja obsługuje wyświetlanie trajektorii wnioskowania, narzędzi po stronie serwera, API OpenAI Responses oraz domyślnie wykorzystuje model GPT-5.6 Luna.","category":"产品更新","source":"Simon Willison 博客","aggregationSource":"Simon Willison 博客","pageTitle":"Wydany LLM 0.32: dodane trajektorie wnioskowania, OpenAI Responses, narzędzia po stronie serwera oraz inteligentniejsze logowanie - Aioga Wiadomości AI","description":"Simon Willison wydał LLM 0.32, najważniejszą nową wersję od czasu uruchomienia projektu. Nowa wersja obsługuje wyświetlanie trajektorii wnioskowania, narzędzi po stronie serwera, A...","url":"https://www.aioga.com/pl/news/cmsfdkkcs1ocxro2ed4z3s131/","contentTranslated":true,"sourceHash":"a2973b505f038783","translatedAt":"2026-08-05T01:23:47.308Z","articleBody":["Dziś rano opublikowałem LLM 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32, jest to najważniejsza nowa wersja LLM od czasu pierwszej publikacji projektu. Nowa wersja obejmuje wsparcie dla widocznych ścieżek wnioskowania, narzędzia dla dostawców po stronie serwera, przeprojektowany dziennik SQLite adresowany przez treść, nowe modele oraz nowe funkcje włączone przez OpenAI Responses API. Opublikowałem również nową wersję wtyczki llm-anthropic: https://github.com/simonw/llm-anthropic, która również zawiera wiele aktualizacji.","Uruchamianie LLM na modelu wnioskowania teraz spowoduje wyświetlenie ich ścieżek wnioskowania w standardowym wyjściu błędów, dzięki czemu możesz zobaczyć ich „myśli”, a te informacje nie będą zawarte w standardowym wyjściu, które możesz przesłać do innych narzędzi. Użycie -R/--hide-reasoning pozwala wyłączyć tę funkcję.","LLM obsługuje modele z serii GPT-5.6 od razu po wyjęciu z pudełka, a nowy domyślny model używający „prompt” w llm to teraz tani, ale potężny GPT-5.6 Luna.","Wywołania LLM mogą teraz korzystać z narzędzi po stronie serwera od różnych dostawców. OpenAI udostępnia środowisko do wykonywania kodu: https://llm.datasette.io/en/stable/openai-models.html#code-interpreter jako narzędzie po stronie serwera; LLM może teraz uruchamiać podpowiedzi korzystające z tego narzędzia, jak pokazano poniżej:","OpenAI zdobyło także narzędzie WebSearch: https://llm.datasette.io/en/stable/openai-models.html#web-search.","llm-anthropic: https://github.com/simonw/llm-anthropic Wtyczka dodała WebSearch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search, WebFetch: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch, CodeExecution: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution oraz AnthropicMCP: https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector, wygląda to tak:","To spowoduje, że Anthropic w pojedynczej interakcji żądanie/odpowiedź ze swoim API wykona wywołanie MCP dla mojego nowego wtyczki datasette-mcp: https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp.","Nowe polecenie punktu końcowego llm OpenAI oferuje narzędzie, które pozwala wykonywać polecenia w dowolnym punkcie końcowym zgodnym z OpenAI: https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it, wystarczy jedno polecenie. Te operacje nie są rejestrowane, więc jest to wygodne narzędzie do wykonywania jednorazowych poleceń dla dowolnego obiektu używającego uniwersalnego języka LLM API.","Oto, jak z niego korzystam, używając uvx (bez potrzeby instalowania LLM) i łącząc z wtyczką narzędzi llm-tools-quickjs: https://github.com/simonw/llm-tools-quickjs, w celu wysyłania poleceń do Gemma 4 12B działającej na moim lokalnym hoście LM Studio: https://lmstudio.ai","Python API LLM wcześniej wymagało utworzenia sesji, a następnie wysyłania jednej wiadomości na raz. To jest abstrakcja prawdziwych cech LLM, każdy żądanie niesie pełną historię wszystkich poprzednich wiadomości. Dla niektórych bardziej zaawansowanych przypadków ta abstrakcja zaczynała stanowić przeszkodę, dlatego nowa wersja wprowadza parametr model.prompt(messages=[]), który można używać w następujący sposób:","Wcześniej LLM zwracały iterowalną sekwencję ciągów znaków dla każdej wskazówki. Działo to dobrze, gdy model zwracał odpowiedzi w formie ciągów znaków, ale nie dało się przewidzieć dziwnych form, które model może wykształcić. Obecnie wiele modeli zwraca mieszane treści, takie jak teksty wniosków, ciągi wyjściowe, wywołania narzędzi, a nawet załączniki obrazów. W LLM 0.32 możesz zamiast tego użyć tej metody: https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events","Łącząc te funkcje, w końcu możemy zaoferować solidną, półstandardową implementację API OpenAI do uzupełniania czatu, którą opublikowałem jako wtyczkę llm-chat-completions-server: https://github.com/simonw/llm-chat-completions-server","Teraz możesz używać nowego polecenia endpoint llm openai przez ten serwer, aby uruchamiać podpowiedzi LLM!","Większym wyzwaniem dla tego typu API jest rejestrowanie logów. Jeśli chcemy wspierać tryb, w którym każdemu żądaniu dodawana jest sekwencja wiadomości, idealnie byłoby unikać zapisywania wszystkich powtarzających się JSON-ów dla każdej sesji.","Rozwiązaniem jest nowy system przechowywania wiadomości adresowalnych według treści: https://llm.datasette.io/en/stable/logging.html#the-message-store, wzorowany na modelowaniu Git. Nowy schemat można zobaczyć w dokumentacji: https://llm.datasette.io/en/stable/logging.html#sql-schema, ale polecenia llm logs i llm logs --json zostały zaktualizowane, aby móc konwertować ten format z powrotem do łatwej w użyciu formy.","To wydanie zawiera również więcej treści. Notatki wydania 0.32: https://llm.datasette.io/en/stable/changelog.html#v0-32 są całkiem obszerne, 0.32rc2: https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30, 0.32rc: https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30, 0.32a3: https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09, 0.32a2: https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12 oraz 0.32a0: https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 mogą uzupełnić wszelkie braki.","Obecne wtyczki LLM powinny nadal działać, ale wtyczki dostarczające dodatkowe modele muszą zostać zaktualizowane do wersji 0.32, aby w pełni uczestniczyć w nowym systemie zdarzeń strumieniowych. W dokumentacji znajduje się przewodnik dotyczący używania uporządkowanych wiadomości i zdarzeń strumieniowych do implementacji wtyczek: https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events.","Zaktualizowałem kilka swoich własnych wtyczek:","Wiele zmian w narzędziach podstawowych w tej wersji zostało napędzanych przez potrzeby Datasette Agent: https://agent.datasette.io/. Kiedy zaczynałem rozwijać LLM, termin „agent” był bardzo niejasno zdefiniowany i odmówiłem jego używania. Do września 2025: https://simonwillison.net/2025/Sep/18/agents/ zaakceptowałem pogląd, że „agenci LLM działają poprzez cykliczne uruchamianie narzędzi w celu osiągnięcia celu”, co było wystarczająco dojrzałe, żebym mógł całkowicie przestać unikać tego terminu.","Łańcuch narzędzi może teraz wstrzymać działanie w oczekiwaniu na zatwierdzenie przez człowieka: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause, i wznowić z zapisanej historii wiadomości: https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume — obie te funkcje są wymagane przez agenta Datasette.","Dziś patrząc na LLM, zaczyna on w moim odczuciu przybierać kształt bardzo przypominający agenta. Jest jedna bardzo ciekawa rzecz: możesz użyć narzędzia wiersza poleceń, aby mieszać i dopasowywać różne narzędzia pochodzące z różnych źródeł i używające różnych modeli, można to zrobić jedną linią kodu, a do tego zawiera wystarczająco potężną bibliotekę Pythona, która pozwala budować systemy takie jak Datasette Agent (https://agent.datasette.io/) i llm-coding-agent (https://github.com/simonw/llm-coding-agent).","Być może następna wersja LLM bezpośrednio włączy koncepcję „agenta” do podstawowej biblioteki. Wciąż próbuję zrozumieć, jak by to wyglądało.","To jest nowa wersja LLM, wydana przez Simona Willisona 4 sierpnia 2026: dodano obsługę śledzenia wnioskowania, odpowiedzi OpenAI, narzędzi po stronie serwera oraz bardziej inteligentnego rejestrowania logów: /2026/Aug/4/.","Część serii artykułów „Nowe wydania dużych modeli językowych (LLM)”: /series/llm-releases/","Poprzedni artykuł: Bezstanowy MCP ponownie wzbudził moje zainteresowanie (i zainspirował mcp-explorer oraz datasette-mcp): /2026/Jul/31/stateless-mcp/","Wspieraj mnie miesięcznie kwotą 10 USD, aby otrzymywać wybrane streszczenia mailowe, podsumowujące najważniejsze wydarzenia związane z LLM w danym miesiącu."],"bodyTranslated":true,"bodySourceHash":"180cd773be407152","bodyTranslatedAt":"2026-08-08T12:28:30.725Z"}}}}