{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-07-23T06:40:50.084Z","headline":"Bonsai 27B：首款可在手机上运行的27B级多模态模型","description":"Bonsai 27B 基于 Qwen3.6 27B，提供三元（1.71 有效比特/权重，5.9 GB）和 1-bit（1.125 有效比特/权重，3.9 GB）两个变体，后者首次将 27B 级模型装入 iPhone 17 Pro。模型支持多步推理、结构化工具调用、视觉任务和计算机使用智能体循环，拥有 262K token 上下文窗口，支持推测解码加速。在 15 项基准测试中，三元变体保留全精度基线 95% 的性能，1-bit 变体保留 90%，数学和编码能力几乎无损。采用 Apache 2.0 许可证开源。","url":"https://www.aioga.com/news/cmrl2m4yg00awbicqo2arojvd/","mainEntityOfPage":"https://www.aioga.com/news/cmrl2m4yg00awbicqo2arojvd/","datePublished":"2026-07-14T19:51:19.560Z","dateModified":"2026-07-14T19:51:19.560Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://prismml.com/news/bonsai-27b","https://aihot.virxact.com/items/cmrl2m4yg00awbicqo2arojvd"],"canonicalUrl":"https://www.aioga.com/news/cmrl2m4yg00awbicqo2arojvd/","directAnswer":{"@type":"Answer","text":"Bonsai 27B 基于 Qwen3.6 27B，提供5.9 GB三元版与3.9 GB的1-bit版。材料称后者可装入 iPhone 17 Pro，并支持多模态、工具调用、262K上下文及推测解码。","url":"https://www.aioga.com/news/cmrl2m4yg00awbicqo2arojvd/","dateCreated":"2026-07-14T19:51:19.560Z","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":"prismml.com source article","url":"https://prismml.com/news/bonsai-27b","datePublished":"2026-07-14T19:51:19.560Z","provider":{"@type":"Organization","name":"prismml.com","url":"https://prismml.com/news/bonsai-27b"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.virxact.com/items/cmrl2m4yg00awbicqo2arojvd","datePublished":"2026-07-14T19:51:19.560Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.virxact.com/items/cmrl2m4yg00awbicqo2arojvd"}}],"aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","originalPublisher":{"name":"prismml.com","url":"https://prismml.com/news/bonsai-27b"},"article":{"id":"cmrl2m4yg00awbicqo2arojvd","slug":"cmrl2m4yg00awbicqo2arojvd","url":"https://www.aioga.com/news/cmrl2m4yg00awbicqo2arojvd/","title":"Bonsai 27B：首款可在手机上运行的27B级多模态模型","title_en":"Bonsai 27B：首款可在手机上运行的27B级模型","summary":"Bonsai 27B 基于 Qwen3.6 27B，提供三元（1.71 有效比特/权重，5.9 GB）和 1-bit（1.125 有效比特/权重，3.9 GB）两个变体，后者首次将 27B 级模型装入 iPhone 17 Pro。模型支持多步推理、结构化工具调用、视觉任务和计算机使用智能体循环，拥有 262K token 上下文窗口，支持推测解码加速。在 15 项基准测试中，三元变体保留全精度基线 95% 的性能，1-bit 变体保留 90%，数学和编码能力几乎无损。采用 Apache 2.0 许可证开源。","source":"Hacker News 热门（buzzing.cc 中文翻译）","sourceUrl":"https://prismml.com/news/bonsai-27b","aiHotUrl":"https://aihot.virxact.com/items/cmrl2m4yg00awbicqo2arojvd","publishedAt":"2026-07-14T19:51:19.560Z","category":"模型更新","score":76,"selected":true,"articleBody":["Today, we're announcing Bonsai 27B, based on Qwen3.6 27B, the new multimodal flagship of the Bonsai family and the first model of its capability class to run on a phone.","Our earlier releases proved that models with 1-bit and ternary weights could produce commercially useful language models. Bonsai 27B extends that frontier to a new capability tier: multi-step reasoning, structured tool calls, vision tasks, and computer-use agentic loops that stay coherent across many steps. Until today, deploying that tier locally has been impractical for a concrete reason: a 27B model occupies roughly 54GB in 16-bit precision, and even a good 4-bit build, at 18GB, is too large for a phone and for most laptops.","Bonsai 27B changes that. It comes in two variants:","Ternary Bonsai 27B uses ternary {−1, 0, +1} weights with FP16 group-wise scaling, giving a true 1.71 effective bits per weight. At 5.9 GB, it is the quality-oriented variant: it runs on an everyday laptop with the full reasoning, tool-calling, and agentic capability.","1-bit Bonsai 27B uses binary {−1, +1} weights with the same group-wise scaling, giving 1.125 effective bits per weight. At 3.9 GB, it is the footprint-oriented variant, which fits within the memory budget of an iPhone 17 Pro, bringing a 27B-class model onto a phone for the first time.","As with every Bonsai release, the low-bit representation runs end to end across the language network, embeddings, attention, MLPs, and the LM head, with no higher-precision escape hatches. Both variants are multimodal, with the vision tower shipping in a compact 4-bit form so on-device workflows can see screenshots, documents, and camera input, not just text. Bonsai 27B carries a full 262K-token context, and supports speculative-decoding, compounding the speed with lossless draft-and-verify acceleration. Everything is available today under the Apache 2.0 License.","Across a 15-benchmark suite spanning knowledge, reasoning, math, coding, instruction following, tool calling, and vision (evaluated in thinking mode, where the model's full reasoning is exercised) Ternary Bonsai 27B retains 95% of the full-precision baseline, and 1-bit Bonsai 27B retains 90% .","Read the table by capability and the story is sharper than the averages: math and coding are nearly untouched, tool calling stays within a few points of full precision - exactly the capabilities that agentic workloads depend on. For comparison, the most aggressive conventional low-bit build of the same base model scores significantly lower than 1-bit Bonsai 27B while occupying 2.5x more memory.","This is the same Pareto shift we demonstrated with our earlier language and image models, now at 27B scale: 27B-class capability at a footprint smaller than a full-precision 2B model. By intelligence density — the measure we introduced with 1-bit Bonsai 8B — 1-bit Bonsai 27B delivers 0.53 per GB: more than 10x the full-precision baseline, and roughly 2.7x the best low-bit alternative available.","The most valuable AI workloads are shifting from single responses to sustained work: assistants that operate real tools, workflows that run unattended before returning a result, and research that synthesizes dozens of documents. This shift changes the shape of the workload — an agent doesn't make one model call, it makes hundreds, each one carrying context, producing structured output, and feeding the next.","Cloud APIs will remain the right choice for many products. But for agentic workloads, cloud-only execution imposes structural constraints: every step is a remote request, per-token cost accumulates with every iteration, and every plan, tool call, and intermediate result crosses the network including the user's private files, screen, and data.","Local execution changes the equation. When a model capable of sustained agentic work fits on the device, the agent can live inside the product: the marginal cost of a hundred-step loop is zero, and the user's data never leaves the machine. Entire categories open up — persistent on-device agents, assistants that work offline, assistants that reason over private local data by construction. What has been missing is a model small enough to deploy this way and capable enough to trust with the work. Bonsai 27B is that model.","Bonsai 27B reaches up to 163 tok/s in 1-bit and 134 tok/s in Ternary on an NVIDIA GeForce RTX 5090. On an M5 Max, it reaches up to 87 tok/s in 1-bit and 58 tok/s in Ternary.","Fitting a phone is a stricter gate than storage numbers suggest. A phone never exposes its full memory to an app - a 12 GB iPhone offers about 6 GB for the model to use on-device, and the model shares that budget with its KV cache and activations. No conventional build of a 27B model comes close to clearing it. At about 4 GB, 1-bit Bonsai 27B is the first to pass through with room to work.","That constraint is why the family ships two deliberate operating points, specifically keeping that in mind: ternary for laptop-class quality, 1-bit for phone-class footprint.","Every Bonsai release has moved the intelligence-per-gigabyte frontier left, and Bonsai 27B moves it past a practical threshold: the full capability set of a modern model with thinking, multimodal understanding, vision, reliable tool use, now fits on the devices people already own.","We believe intelligence density will be one of the defining axes of the next stage of AI progress. Raw capability determines what a model can do; density determines where it can do it. Every leftward shift of the frontier expands the set of devices, products, and environments where advanced AI can operate and changes the economics of every deployment surface it touches, from phones to single-GPU serving. The methodology behind Bonsai is architecture-agnostic, and the frontier will keep moving: larger models and new architectures are already in progress.","Early computers filled rooms; today they live in our pockets. Intelligence is making the same journey, and Bonsai 27B is its largest step yet. Platform Coverage","Bonsai 27B runs natively on Apple devices (Mac, iPhone, iPad) via MLX and on NVIDIA GPUs via CUDA, through custom low-bit kernels built for its hybrid-attention architecture. Model weights are available today under the Apache 2.0 License. With this release, we’re offering a free, limited-time developer preview API so developers can easily try our model.","Full technical details of our compression, evaluation, and benchmarking processes are available in our whitepaper：https://github.com/PrismML-Eng/Bonsai-demo/blob/main/bonsai-27b-whitepaper.pdf.","PrismML emerged from a team of Caltech researchers and was founded with support from Khosla Ventures, Cerberus, and Google, with continuing support from Samsung. We've spent years tackling one of the field's hardest problems: compressing neural networks without sacrificing their reasoning ability.","If you want to help build the next generation of state-of-the-art AI, we'd love to hear from you. Check out our careers page：/careers."],"articleImages":[{"sourceUrl":"https://cdn.prod.website-files.com/699604cc2b9dd89bdbda0608/6a564448065275919f2414d0_id.webp","alt":"","afterParagraph":8,"url":"/media/articles/cmrl2m4yg00awbicqo2arojvd/516cfaf6dc366444.webp"},{"sourceUrl":"https://cdn.prod.website-files.com/697a3312d33c2cc715ec3899/69cac0e755d4efa49449bebe_hf-logo-monochrome%201.svg","alt":"","afterParagraph":21,"url":"/media/articles/cmrl2m4yg00awbicqo2arojvd/64e84a7dc1bb0768.jpg"}],"mediaStatus":"ok","articleBodyZh":["今天，我们宣布推出 Bonsai 27B，它基于 Qwen3.6 27B，是 Bonsai 系列的新多模态旗舰，也是首款可在手机上运行的同类能力模型。","我们之前的版本证明，使用 1-bit 和三值权重的模型也可以产生商业上有用的语言模型。Bonsai 27B 将这一前沿扩展到新的能力等级：多步推理、结构化工具调用、视觉任务以及可在多步中保持连贯性的计算机使用型代理循环。在今天之前，本地部署这一能力层是不切实际的，原因很具体：27B 模型在 16-bit 精度下大约占用 54GB，即使是较好的 4-bit 构建，18GB 的大小也对于手机及大多数笔记本电脑来说过大。","Bonsai 27B 改变了这一局面。它有两个变体：","三值 Bonsai 27B 使用三值 {−1, 0, 1} 权重和 FP16 分组缩放，每个权重的有效位数为 1.71 位。其大小为 5.9 GB，是面向质量的变体：它可以在普通笔记本电脑上运行，具备完整的推理、工具调用和代理能力。","1-bit Bonsai 27B 使用二值 {−1, 1} 权重，并使用相同的分组缩放，每个权重的有效位数为 1.125 位。其大小为 3.9 GB，是面向占用空间的变体，可在 iPhone 17 Pro 的内存预算内运行，使 27B 级模型首次能在手机上使用。","与每一次 Bonsai 发布一样，这种低位表示运行覆盖语言网络、嵌入、注意力、MLP 和语言模型头部的端到端过程，不依赖更高精度的逃逸路径。两个变体都是多模态的，视觉模块以紧凑的 4-bit 形式提供，使设备端工作流能够识别屏幕截图、文档和摄像头输入，而不仅仅是文本。Bonsai 27B 支持完整的 262K 令牌上下文，并支持预测式解码，通过无损的草稿-验证加速来提高速度。所有内容今天都已在 Apache 2.0 许可下提供。","在一个涵盖知识、推理、数学、编程、指令遵循、工具调用和视觉的 15 个基准测试套件中（在“思考模式”下评估，即模型的全部推理能力被使用），三值 Bonsai 27B 保留了 95% 的全精度基线性能，而 1-bit Bonsai 27B 保留了 90% 的性能。","通过能力读取表格，故事比平均水平更清晰：数学和编码几乎未受影响，工具调用保持在接近完全精度的几个点内——正是智能工作负载依赖的能力。作为对比，同一基础模型最激进的常规模型低位构建，其得分显著低于1-bit Bonsai 27B，同时占用的内存多2.5倍。","这与我们在早期语言和图像模型中展示的帕累托变化相同，现在在27B规模下实现：在占用比全精度2B模型更小的内存下实现27B级别的能力。按照智能密度——我们在1-bit Bonsai 8B中引入的衡量标准——1-bit Bonsai 27B每GB提供0.53的能力：是全精度基线的10倍以上，大约是现有最佳低位替代品的2.7倍。","最有价值的AI工作负载正在从单次响应转向持续工作：操作真实工具的助手、在无人干预下运行工作流程然后返回结果的系统，以及汇总几十份文档的研究。这个变化改变了工作负载的形态——一个智能体不只调用一次模型，而是调用数百次，每次都携带上下文，生成结构化输出，并用于下一步操作。","云API仍将是许多产品的正确选择。但对于智能工作负载，仅靠云执行会带来结构性限制：每一步都是远程请求，每个标记的成本随着每次迭代累积，每个计划、工具调用和中间结果都需通过网络，包括用户的私有文件、屏幕和数据。","本地执行改变了这一计算方式。当一个能够持续进行智能工作的模型能够适配到设备上时，智能体可以运行在产品内部：百步循环的边际成本为零，而且用户的数据永远不会离开设备。这开辟了全新的类别——持久化本地智能体、离线工作的助手、能够基于本地私有数据进行推理的助手。此前缺少的，是一个既小巧到可以这样部署，又足够强大可以信赖其处理工作的模型。Bonsai 27B正是这样的模型。","Bonsai 27B在NVIDIA GeForce RTX 5090上，1-bit模式下速度可达每秒163个标记，Ternary模式下为每秒134个标记。在M5 Max上，1-bit模式下速度可达每秒87个标记，Ternary模式下为每秒58个标记。","在手机上部署模型比存储容量数字显示的更严格。手机不会将其全部内存暴露给应用——一部 12 GB 的 iPhone 大约只提供 6 GB 给模型在设备上使用，而且模型还要与其 KV 缓存和激活共享这部分内存。没有常规构建的 27B 模型能接近清空这部分内存。约 4 GB 时，1-bit Bonsai 27B 是第一个能够通过并有操作空间的模型。","正因为这个限制，这个系列提供了两个特意设计的操作点，特别考虑到了这一点：三值（ternary）用于笔记本级质量，1-bit 用于手机级占用。","每一个 Bonsai 版本都推动了每 GB 智能密度的边界，而 Bonsai 27B 将其推过了一个实际阈值：现代模型的全部能力集，包括思考、多模态理解、视觉、可靠的工具使用，现在都能适配人们已有的设备上。","我们相信智能密度将成为下一阶段 AI 进展的决定性轴之一。原始能力决定模型能做什么；密度决定它能在哪做。边界的每一次左移都会扩大高级 AI 可运行的设备、产品和环境，并改变它触及的每一个部署场景的经济效益，从手机到单 GPU 服务节点。Bonsai 的方法论与架构无关，边界将继续推进：更大模型和新架构正在开发中。","早期计算机占据整个房间；今天它们存在于我们的口袋里。智能正在经历同样的旅程，而 Bonsai 27B 是迄今为止最大的跨步。平台覆盖","Bonsai 27B 可以通过 MLX 在 Apple 设备（Mac、iPhone、iPad）上原生运行，也可以通过 CUDA 在 NVIDIA GPU 上运行，通过为其混合注意力架构定制的低比特核实现。模型权重今日即可在 Apache 2.0 许可证下获取。基于此次发布，我们提供免费的限时开发者预览 API，方便开发者轻松试用我们的模型。","我们关于压缩、评估和基准测试流程的完整技术细节可参考白皮书：https://github.com/PrismML-Eng/Bonsai-demo/blob/main/bonsai-27b-whitepaper.pdf","PrismML 起源于一支加州理工学院的研究团队，并在 Khosla Ventures、Cerberus 和 Google 的支持下成立，同时持续获得三星的支持。我们花了多年时间攻克该领域最难的问题之一：在不牺牲神经网络推理能力的前提下进行压缩。","如果你想帮助构建下一代最先进的人工智能，我们非常希望听到你的声音。请查看我们的招聘页面：/careers。"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"Bonsai 27B 基于 Qwen3.6 27B，提供5.9 GB三元版与3.9 GB的1-bit版。材料称后者可装入 iPhone 17 Pro，并支持多模态、工具调用、262K上下文及推测解码。","background":"公开材料称，27B模型以16位精度部署约需54 GB，常规4位构建仍约18 GB。Bonsai通过端到端低比特权重压缩语言网络，并将视觉塔以紧凑的4位形式发布。","viewpoint":"Aioga 判断，此次更新的核心价值不是单项能力增加，而是把27B级多模态模型的本地部署体积降至手机和日常笔记本可承载的范围，同时尽量维持推理与智能体相关能力。","implications":"这可能扩大端侧文档、截图和相机输入处理的可行空间，并降低部分工作流对远程推理的依赖。值得关注的是，材料中的性能结论来自15项基准，尚不能替代真实设备体验。","nextStep":"建议后续核验公开模型、许可证与运行代码，并在实际设备上比较内存占用、速度、能耗和长上下文稳定性；同时复测工具调用、视觉任务以及多步智能体循环的完成质量。","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":"2026-07-22T19:31:35.456Z","sourceHash":"1fbd6b4efbb143fe","review":{"approved":true,"groundedness":96,"clarity":92,"duplicationRisk":18,"blockingIssues":[],"notes":["候选内容对关键参数、功能、部署体积和基准测试范围的表述均可由来源材料支持。","将性能结论限定为材料中的15项基准，并提示仍需真实设备验证，避免了把发布方测试结果外推为实际体验。","“常规4位构建”可更精确地写成来源所称的“较好的4位构建”，但这属于可选措辞优化，不影响事实成立。"]},"validation":{"passed":true,"mode":"ai-auto","revisions":0,"checks":["schema","length","source-attribution","low-source-overlap","no-html","independent-ai-review"]}},"tags":["模型更新","Hacker News 热门（buzzing.cc 中文翻译）"],"translations":{"zh-CN":{"title":"Bonsai 27B：首款可在手机上运行的27B级多模态模型","summary":"Bonsai 27B 基于 Qwen3.6 27B，提供三元（1.71 有效比特/权重，5.9 GB）和 1-bit（1.125 有效比特/权重，3.9 GB）两个变体，后者首次将 27B 级模型装入 iPhone 17 Pro。模型支持多步推理、结构化工具调用、视觉任务和计算机使用智能体循环，拥有 262K token 上下文窗口，支持推测解码加速。在 15 项基准测试中，三元变体保留全精度基线 95% 的性能，1-bit 变体保留 90%，数学和编码能力几乎无损。采用 Apache 2.0 许可证开源。","category":"模型更新","source":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Bonsai 27B：首款可在手机上运行的27B级多模态模型 - Aioga AI资讯","description":"Bonsai 27B 基于 Qwen3.6 27B，提供三元（1.71 有效比特/权重，5.9 GB）和 1-bit（1.125 有效比特/权重，3.9 GB）两个变体，后者首次将 27B 级模型装入 iPhone 17 Pro。模型支持多步推理、结构化工具调用、视觉任务和计算机使用智能体循环，拥有 262K token 上下文窗口，支持推测解码加速。在 1","url":"https://www.aioga.com/news/cmrl2m4yg00awbicqo2arojvd/"},"en":{"title":"Bonsai 27B: The first 27B-level multimodal model to run on a smartphone","summary":"The Bonsai 27B is based on Qwen 3.6 27B, offering two variants: ternary (1.71 effective bits/weight, 5.9 GB) and 1-bit (1.125 effective bits/weight, 3.9 GB), the latter of which is the first to bring a 27B-level model into the iPhone 17 Pro. The model supports multi-step inference, structured tool calls, vision tasks, and agent loops for computer use, with a 262K token context window and accelerated inference decoding. In 15 benchmarks, the ternary variant retained 95% of the full-precision baseline performance, while the 1-bit variant retained 90%, with nearly no loss of mathematical and coding capabilities. Open source under the Apache 2.0 license.","category":"Models","source":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Bonsai 27B: The first 27B-level multimodal model to run on a smartphone - Aioga AI News","description":"The Bonsai 27B is based on Qwen 3.6 27B, offering two variants: ternary (1.71 effective bits/weight, 5.9 GB) and 1-bit (1.125 effective bits/weight, 3.9 GB), the latter of which is","url":"https://www.aioga.com/en/news/cmrl2m4yg00awbicqo2arojvd/","contentTranslated":true,"sourceHash":"40b647793fcd518e","translatedAt":"2026-07-19T12:15:04.641Z"},"ja":{"title":"盆栽27B:スマートフォンで動作する初の27Bレベルのマルチモーダルモデル","summary":"Bonsai 27BはQwen 3.6 27Bをベースにしており、2つのバリアントがあります:三進法(1.71実効ビット/重量、5.9GB)と1ビット(1.125実効ビット/重量、3.9GB)です。後者はiPhone 17 Proに初めて27Bレベルのモデルを搭載しました。 このモデルは、多段階推論、構造化ツール呼び出し、ビジョンタスク、エージェントループをコンピュータ用途でサポートし、262Kトークンコンテキストウィンドウと加速推論デコードを備えています。 15のベンチマークでは、三進バリアントはフルクシオのベースライン性能の95%を維持し、1ビットバリアントは90%を維持し、数学的およびコーディング能力の損失はほとんどありませんでした。 Apache 2.0ライセンスのもとでオープンソース。","category":"モデル更新","source":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"盆栽27B:スマートフォンで動作する初の27Bレベルのマルチモーダルモデル - Aioga AIニュース","description":"Bonsai 27BはQwen 3.6 27Bをベースにしており、2つのバリアントがあります:三進法(1.71実効ビット/重量、5.9GB)と1ビット(1.125実効ビット/重量、3.9GB)です。後者はiPhone 17 Proに初めて27Bレベルのモデルを搭載しました。 このモデルは、多段階推論、構造化ツール呼び出し、ビジョンタスク、エージェントループを","url":"https://www.aioga.com/ja/news/cmrl2m4yg00awbicqo2arojvd/","contentTranslated":true,"sourceHash":"40b647793fcd518e","translatedAt":"2026-07-19T12:15:04.798Z"},"ko":{"title":"분재 27B: 스마트폰에서 구동되는 최초의 27B 레벨 멀티모달 모델","summary":"분사이 27B는 Qwen 3.6 27B를 기반으로 하며, 두 가지 변형을 제공합니다: 삼진형(1.71 유효 비트/가중치, 5.9 GB)과 1비트(1.125 유효 비트/가중치, 3.9 GB)입니다. 후자는 아이폰 17 프로에 27B 등급 모델을 처음으로 도입한 모델입니다. 이 모델은 컴퓨터용 다단계 추론, 구조화된 도구 호출, 비전 작업, 에이전트 루프를 지원하며, 262K 토큰 컨텍스트 창과 가속화된 추론 디코딩을 지원합니다. 15개의 벤치마크에서 삼진법 변형은 전정밀도 기준선 성능의 95%를 유지했으며, 1비트 변형은 90%를 유지했으며 수학 및 코딩 능력의 손실은 거의 없었습니다. Apache 2.0 라이선스 하에 오픈 소스입니다.","category":"모델 업데이트","source":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"분재 27B: 스마트폰에서 구동되는 최초의 27B 레벨 멀티모달 모델 - Aioga AI 뉴스","description":"분사이 27B는 Qwen 3.6 27B를 기반으로 하며, 두 가지 변형을 제공합니다: 삼진형(1.71 유효 비트/가중치, 5.9 GB)과 1비트(1.125 유효 비트/가중치, 3.9 GB)입니다. 후자는 아이폰 17 프로에 27B 등급 모델을 처음으로 도입한 모델입니다. 이 모델은 컴퓨터용 다단계 추론, 구조화된 도구 호","url":"https://www.aioga.com/ko/news/cmrl2m4yg00awbicqo2arojvd/","contentTranslated":true,"sourceHash":"40b647793fcd518e","translatedAt":"2026-07-19T12:15:04.709Z"},"es":{"title":"Bonsai 27B: El primer modelo multimodal de nivel 27B que funciona en un smartphone","summary":"El Bonsai 27B está basado en el Qwen 3.6 27B, ofreciendo dos variantes: ternario (1,71 bits efectivos/peso, 5,9 GB) y 1 bit (1,125 bits/peso, 3,9 GB), esta última siendo la primera en incorporar un modelo de nivel 27B al iPhone 17 Pro. El modelo soporta inferencia en varios pasos, llamadas estructuradas a herramientas, tareas de visión y bucles de agentes para uso informático, con una ventana de contexto de 262K tokens y decodificación acelerada por inferencia. En 15 benchmarks, la variante ternaria mantuvo el 95% del rendimiento base de precisión completa, mientras que la variante de 1 bit conservó el 90%, con casi ninguna pérdida de capacidades matemáticas y de codificación. Código abierto bajo la licencia Apache 2.0.","category":"Modelos","source":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Bonsai 27B: El primer modelo multimodal de nivel 27B que funciona en un smartphone - Aioga Noticias de IA","description":"El Bonsai 27B está basado en el Qwen 3.6 27B, ofreciendo dos variantes: ternario (1,71 bits efectivos/peso, 5,9 GB) y 1 bit (1,125 bits/peso, 3,9 GB), esta última siendo la primera","url":"https://www.aioga.com/es/news/cmrl2m4yg00awbicqo2arojvd/","contentTranslated":true,"sourceHash":"40b647793fcd518e","translatedAt":"2026-07-19T12:15:04.880Z"},"fr":{"title":"Bonsai 27B : Le premier modèle multimodal de niveau 27B à fonctionner sur un smartphone","summary":"Le Bonsai 27B est basé sur le Qwen 3.6 27B, proposant deux variantes : ternaire (1,71 bits effectifs/poids, 5,9 Go) et 1 bit (1,125 bits/poids, 3,9 Go), cette dernière étant la première à intégrer un modèle de niveau 27B sur l’iPhone 17 Pro. Le modèle prend en compte l’inférence en plusieurs étapes, les appels d’outils structurés, les tâches de vision et les boucles d’agents pour l’utilisation informatique, avec une fenêtre de contexte de 262K jetons et un décodage accéléré par inférence. En 15 benchmarks, la variante ternaire conservait 95 % des performances de base en pleine précision, tandis que la variante 1 bit conservait 90 %, avec presque aucune perte de capacités mathématiques et de codage. Open source sous licence Apache 2.0.","category":"Modèles","source":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Bonsai 27B : Le premier modèle multimodal de niveau 27B à fonctionner sur un smartphone - Aioga Actualités IA","description":"Le Bonsai 27B est basé sur le Qwen 3.6 27B, proposant deux variantes : ternaire (1,71 bits effectifs/poids, 5,9 Go) et 1 bit (1,125 bits/poids, 3,9 Go), cette dernière étant la pre","url":"https://www.aioga.com/fr/news/cmrl2m4yg00awbicqo2arojvd/","contentTranslated":true,"sourceHash":"40b647793fcd518e","translatedAt":"2026-07-19T12:15:04.834Z"},"de":{"title":"Bonsai 27B: Das erste multimodale Modell auf 27B-Niveau, das auf einem Smartphone läuft","summary":"Der Bonsai 27B basiert auf Qwen 3.6 27B und bietet zwei Varianten an: Ternary (1,71 effektive Bits/Gewicht, 5,9 GB) und 1-Bit (1.125 effektive Bits/Gewicht, 3,9 GB), wobei letzterer der erste ist, der ein 27B-Modell ins iPhone 17 Pro einführt. Das Modell unterstützt mehrstufige Inferenz, strukturierte Werkzeugaufrufe, Vision-Aufgaben und Agentenschleifen für den Computergebrauch, mit einem 262K-Token-Kontextfenster und beschleunigter Inferenzdekodierung. In 15 Benchmarks behielt die ternäre Variante 95 % der Vollpräzisions-Basisleistung, während die 1-Bit-Variante 90 % behielt, mit nahezu keinem Verlust an mathematischen und kodierbaren Fähigkeiten. Open Source unter der Apache 2.0-Lizenz.","category":"模型更新","source":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Bonsai 27B: Das erste multimodale Modell auf 27B-Niveau, das auf einem Smartphone läuft - Aioga KI-News","description":"Der Bonsai 27B basiert auf Qwen 3.6 27B und bietet zwei Varianten an: Ternary (1,71 effektive Bits/Gewicht, 5,9 GB) und 1-Bit (1.125 effektive Bits/Gewicht, 3,9 GB), wobei letztere","url":"https://www.aioga.com/de/news/cmrl2m4yg00awbicqo2arojvd/","contentTranslated":true,"sourceHash":"40b647793fcd518e","translatedAt":"2026-07-19T12:15:11.141Z"},"pt-BR":{"title":"Bonsai 27B: O primeiro modelo multimodal nível 27B a rodar em um smartphone","summary":"O Bonsai 27B é baseado no Qwen 3.6 27B, oferecendo duas variantes: ternária (1,71 bits efetivos/peso, 5,9 GB) e 1 bit (1,125 bits/peso, 3,9 GB), sendo esta última a trazer um modelo nível 27B para o iPhone 17 Pro. O modelo suporta inferência em múltiplas etapas, chamadas estruturadas de ferramentas, tarefas de visão e loops de agentes para uso em computador, com uma janela de contexto de 262K tokens e decodificação acelerada de inferência. Em 15 benchmarks, a variante ternária manteve 95% do desempenho base de precisão total, enquanto a variante de 1 bit manteve 90%, com quase nenhuma perda de capacidades matemáticas e de codificação. Código aberto sob a licença Apache 2.0.","category":"模型更新","source":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Bonsai 27B: O primeiro modelo multimodal nível 27B a rodar em um smartphone - Aioga Notícias de IA","description":"O Bonsai 27B é baseado no Qwen 3.6 27B, oferecendo duas variantes: ternária (1,71 bits efetivos/peso, 5,9 GB) e 1 bit (1,125 bits/peso, 3,9 GB), sendo esta última a trazer um model","url":"https://www.aioga.com/pt-BR/news/cmrl2m4yg00awbicqo2arojvd/","contentTranslated":true,"sourceHash":"40b647793fcd518e","translatedAt":"2026-07-19T12:15:09.313Z"},"ru":{"title":"Bonsai 27B: первая мультимодальная модель уровня 27B, работающая на смартфоне","summary":"Bonsai 27B основан на Qwen 3.6 27B и предлагает два варианта: трёхкратный (1,71 эффективных бит/вес, 5,9 ГБ) и 1-битный (1,125 эффективных бит/вес, 3,9 ГБ), последний из которых первым внедряет модель уровня 27B в iPhone 17 Pro. Модель поддерживает многоступенчатый вывод, структурированные вызовы инструментов, задачи зрения и циклы агентов для компьютерного использования, с окном контекста токена 262K и ускоренным декодированием вывода. В 15 бенчмарках третичный вариант сохранил 95% базовой производительности полной точности, тогда как 1-битный вариант — 90%, практически без потери математических и кодовых возможностей. Открытый исходный код под лицензией Apache 2.0.","category":"模型更新","source":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Bonsai 27B: первая мультимодальная модель уровня 27B, работающая на смартфоне - Aioga Новости ИИ","description":"Bonsai 27B основан на Qwen 3.6 27B и предлагает два варианта: трёхкратный (1,71 эффективных бит/вес, 5,9 ГБ) и 1-битный (1,125 эффективных бит/вес, 3,9 ГБ), последний из которых пе","url":"https://www.aioga.com/ru/news/cmrl2m4yg00awbicqo2arojvd/","contentTranslated":true,"sourceHash":"40b647793fcd518e","translatedAt":"2026-07-19T12:15:10.005Z"},"ar":{"title":"بونساي 27B: أول نموذج متعدد الوسائط بمستوى 27B يعمل على هاتف ذكي","summary":"يعتمد بونساي 27B على معالج Qwen 3.6 27B، ويقدم نسختين: ثلاثي (1.71 بت فعال/وزن، 5.9 جيجابايت) و1-بت (1.125 بت فعال/وزن، 3.9 جيجابايت)، والأخير هو الأول الذي يدخل نموذج 27B في آيفون 17 برو. يدعم النموذج الاستدلال متعدد الخطوات، واستدعاءات الأدوات المنظمة، ومهام الرؤية، وحلقات الوكيل للاستخدام الحاسوبي، مع نافذة سياق رمزية بقيمة 262K وفك ترميز استدلالي معجل. في 15 معيارا، احتفظ النسخة الثلاثية بنسبة 95٪ من الأداء الأساسي الكامل للدقة، بينما احتفظ النسخة ذات البت الواحد بنسبة 90٪، دون فقدان شبه كامل للقدرات الرياضية والترميزية. مفتوح المصدر تحت رخصة Apache 2.0.","category":"模型更新","source":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"بونساي 27B: أول نموذج متعدد الوسائط بمستوى 27B يعمل على هاتف ذكي - Aioga أخبار الذكاء الاصطناعي","description":"يعتمد بونساي 27B على معالج Qwen 3.6 27B، ويقدم نسختين: ثلاثي (1.71 بت فعال/وزن، 5.9 جيجابايت) و1-بت (1.125 بت فعال/وزن، 3.9 جيجابايت)، والأخير هو الأول الذي يدخل نموذج 27B في آيفون","url":"https://www.aioga.com/ar/news/cmrl2m4yg00awbicqo2arojvd/","contentTranslated":true,"sourceHash":"40b647793fcd518e","translatedAt":"2026-07-19T12:15:10.899Z"},"hi":{"title":"बोन्साई 27बी: स्मार्टफोन पर चलने वाला पहला 27बी-स्तरीय मल्टीमॉडल मॉडल","summary":"बोन्साई 27बी Qwen 3.6 27B पर आधारित है, जो दो वेरिएंट पेश करता है: टर्नरी (1.71 प्रभावी बिट्स/वजन, 5.9 जीबी) और 1-बिट (1.125 प्रभावी बिट्स/वजन, 3.9 जीबी), जिनमें से बाद वाला iPhone 27 Pro में 17B-स्तरीय मॉडल लाने वाला पहला है। मॉडल 262K टोकन संदर्भ विंडो और त्वरित अनुमान डिकोडिंग के साथ कंप्यूटर उपयोग के लिए बहु-चरणीय अनुमान, संरचित टूल कॉल, दृष्टि कार्यों और एजेंट लूप का समर्थन करता है। 15 बेंचमार्क में, टर्नरी संस्करण ने पूर्ण-सटीक आधारभूत प्रदर्शन का 95% बरकरार रखा, जबकि 1-बिट संस्करण ने 90% को बरकरार रखा, जिसमें गणितीय और कोडिंग क्षमताओं का लगभग कोई नुकसान नहीं हुआ। अपाचे 2.0 लाइसेंस के तहत खुला स्रोत।","category":"模型更新","source":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"बोन्साई 27बी: स्मार्टफोन पर चलने वाला पहला 27बी-स्तरीय मल्टीमॉडल मॉडल - Aioga AI समाचार","description":"बोन्साई 27बी Qwen 3.6 27B पर आधारित है, जो दो वेरिएंट पेश करता है: टर्नरी (1.71 प्रभावी बिट्स/वजन, 5.9 जीबी) और 1-बिट (1.125 प्रभावी बिट्स/वजन, 3.9 जीबी), जिनमें से बाद वाला iPhone","url":"https://www.aioga.com/hi/news/cmrl2m4yg00awbicqo2arojvd/","contentTranslated":true,"sourceHash":"40b647793fcd518e","translatedAt":"2026-07-19T12:15:12.241Z"},"it":{"title":"Bonsai 27B: Il primo modello multimodale di livello 27B a funzionare su uno smartphone","summary":"Il Bonsai 27B si basa sul Qwen 3.6 27B, offrendo due varianti: ternaria (1,71 bit/peso effettivo, 5,9 GB) e 1-bit (1,125 bit/peso effettivo, 3,9 GB), quest'ultima la prima a portare un modello di livello 27B nell'iPhone 17 Pro. Il modello supporta inferenza a più passi, chiamate a strumenti strutturati, task di visione e cicli di agenti per l'uso computerizzato, con una finestra di contesto di token di 262K e decodifica accelerata per inferenza. In 15 benchmark, la variante ternaria ha mantenuto il 95% delle prestazioni di base a piena precisione, mentre la variante a 1 bit ha mantenuto il 90%, con quasi nessuna perdita di capacità matematiche e di codifica. Open source sotto licenza Apache 2.0.","category":"模型更新","source":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Bonsai 27B: Il primo modello multimodale di livello 27B a funzionare su uno smartphone - Aioga Notizie IA","description":"Il Bonsai 27B si basa sul Qwen 3.6 27B, offrendo due varianti: ternaria (1,71 bit/peso effettivo, 5,9 GB) e 1-bit (1,125 bit/peso effettivo, 3,9 GB), quest'ultima la prima a portar","url":"https://www.aioga.com/it/news/cmrl2m4yg00awbicqo2arojvd/","contentTranslated":true,"sourceHash":"40b647793fcd518e","translatedAt":"2026-07-19T12:15:09.986Z"},"nl":{"title":"Bonsai 27B: Het eerste multimodale model op 27B-niveau dat op een smartphone draait","summary":"De Bonsai 27B is gebaseerd op Qwen 3.6 27B en biedt twee varianten: ternaire (1,71 effectieve bits/gewicht, 5,9 GB) en 1-bit (1,125 effectieve bits/gewicht, 3,9 GB), waarvan de laatste de eerste is die een 27B-model in de iPhone 17 Pro brengt. Het model ondersteunt meerstapsinferentie, gestructureerde tooloproepen, visietaken en agentlussen voor computergebruik, met een 262K tokencontextvenster en versnelde inferentiedecodering. In 15 benchmarks behield de ternaire variant 95% van de volledige precisie basisprestaties, terwijl de 1-bit variant 90% behield, met vrijwel geen verlies aan wiskundige en codeermogelijkheden. Open source onder de Apache 2.0-licentie.","category":"模型更新","source":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Bonsai 27B: Het eerste multimodale model op 27B-niveau dat op een smartphone draait - Aioga AI-nieuws","description":"De Bonsai 27B is gebaseerd op Qwen 3.6 27B en biedt twee varianten: ternaire (1,71 effectieve bits/gewicht, 5,9 GB) en 1-bit (1,125 effectieve bits/gewicht, 3,9 GB), waarvan de laa","url":"https://www.aioga.com/nl/news/cmrl2m4yg00awbicqo2arojvd/","contentTranslated":true,"sourceHash":"40b647793fcd518e","translatedAt":"2026-07-19T12:15:09.519Z"},"tr":{"title":"Bonsai 27B: Akıllı telefonda çalışan ilk 27B seviyeli çoklu modal model","summary":"Bonsai 27B, Qwen 3.6 27B temelli olup, iki varyant sunuyor: üçlü (1.71 etkili bit/ağırlık, 5.9 GB) ve 1-bit (1.125 etkili bit/ağırlık, 3.9 GB); ikincisi, iPhone 17 Pro'ya ilk kez 27B seviyesinde bir model getiren modeldir. Model, çok adımlı çıkarım, yapılandırılmış araç çağrıları, vizyon görevleri ve bilgisayar kullanımı için ajan döngülerini destekler; 262K token bağlam penceresi ve hızlandırılmış çıkarım kodlama ile donatılmıştır. 15 benchmark'ta, üçlü varyant tam hassasiyetli temel performansın %95'ini korurken, 1-bit varyant %90'ını korudu ve matematiksel ve kodlama yeteneklerinde neredeyse hiç kaybolmadı. Apache 2.0 lisansı altında açık kaynak.","category":"模型更新","source":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Bonsai 27B: Akıllı telefonda çalışan ilk 27B seviyeli çoklu modal model - Aioga AI Haberleri","description":"Bonsai 27B, Qwen 3.6 27B temelli olup, iki varyant sunuyor: üçlü (1.71 etkili bit/ağırlık, 5.9 GB) ve 1-bit (1.125 etkili bit/ağırlık, 3.9 GB); ikincisi, iPhone 17 Pro'ya ilk kez 2","url":"https://www.aioga.com/tr/news/cmrl2m4yg00awbicqo2arojvd/","contentTranslated":true,"sourceHash":"40b647793fcd518e","translatedAt":"2026-07-19T12:15:11.878Z"},"vi":{"title":"Bonsai 27B: Mô hình đa phương thức cấp 27B đầu tiên chạy trên điện thoại thông minh","summary":"Bonsai 27B dựa trên Qwen 3.6 27B, cung cấp hai biến thể: bậc ba (1.71 bit hiệu dụng / trọng lượng, 5.9 GB) và 1 bit (1.125 bit hiệu dụng / trọng lượng, 3.9 GB), sau này là biến thể đầu tiên đưa mô hình cấp 27B vào iPhone 17 Pro. Mô hình hỗ trợ suy luận nhiều bước, lệnh gọi công cụ có cấu trúc, tác vụ thị giác và vòng lặp tác nhân để sử dụng máy tính, với cửa sổ ngữ cảnh mã thông báo 262K và giải mã suy luận tăng tốc. Trong 15 điểm chuẩn, biến thể bậc ba giữ lại 95% hiệu suất cơ sở có độ chính xác đầy đủ, trong khi biến thể 1 bit giữ lại 90%, gần như không mất khả năng toán học và mã hóa. Mã nguồn mở theo giấy phép Apache 2.0.","category":"模型更新","source":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Bonsai 27B: Mô hình đa phương thức cấp 27B đầu tiên chạy trên điện thoại thông minh - Tin tức AI Aioga","description":"Bonsai 27B dựa trên Qwen 3.6 27B, cung cấp hai biến thể: bậc ba (1.71 bit hiệu dụng / trọng lượng, 5.9 GB) và 1 bit (1.125 bit hiệu dụng / trọng lượng, 3.9 GB), sau này là biến thể","url":"https://www.aioga.com/vi/news/cmrl2m4yg00awbicqo2arojvd/","contentTranslated":true,"sourceHash":"40b647793fcd518e","translatedAt":"2026-07-19T12:15:11.438Z"},"id":{"title":"Bonsai 27B: Model multimoda tingkat 27B pertama yang berjalan di smartphone","summary":"Bonsai 27B didasarkan pada Qwen 3.6 27B, menawarkan dua varian: terner (1.71 bit efektif / berat, 5.9 GB) dan 1-bit (1.125 bit efektif / berat, 3.9 GB), yang terakhir adalah yang pertama membawa model level 27B ke dalam iPhone 17 Pro. Model ini mendukung inferensi multi-langkah, panggilan alat terstruktur, tugas visi, dan loop agen untuk penggunaan komputer, dengan jendela konteks token 262K dan decoding inferensi yang dipercepat. Dalam 15 tolok ukur, varian terner mempertahankan 95% dari kinerja baseline presisi penuh, sedangkan varian 1-bit mempertahankan 90%, dengan hampir tidak kehilangan kemampuan matematika dan pengkodean. Sumber terbuka di bawah lisensi Apache 2.0.","category":"模型更新","source":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Bonsai 27B: Model multimoda tingkat 27B pertama yang berjalan di smartphone - Berita AI Aioga","description":"Bonsai 27B didasarkan pada Qwen 3.6 27B, menawarkan dua varian: terner (1.71 bit efektif / berat, 5.9 GB) dan 1-bit (1.125 bit efektif / berat, 3.9 GB), yang terakhir adalah yang p","url":"https://www.aioga.com/id/news/cmrl2m4yg00awbicqo2arojvd/","contentTranslated":true,"sourceHash":"40b647793fcd518e","translatedAt":"2026-07-19T12:15:09.284Z"},"th":{"title":"Bonsai 27B: โมเดลมัลติโมดอลระดับ 27B รุ่นแรกที่ทํางานบนสมาร์ทโฟน","summary":"Bonsai 27B ใช้ Qwen 3.6 27B ซึ่งมีสองรุ่น: ไตรภาค (1.71 บิตที่ใช้งานจริง/น้ําหนัก, 5.9 GB) และ 1 บิต (1.125 บิตที่ใช้งานจริง/น้ําหนัก, 3.9 GB) ซึ่งเป็นรุ่นแรกที่นํารุ่นระดับ 27B มาสู่ iPhone 17 Pro โมเดลนี้รองรับการอนุมานแบบหลายขั้นตอน การเรียกเครื่องมือที่มีโครงสร้าง งานวิสัยทัศน์ และลูปตัวแทนสําหรับการใช้งานคอมพิวเตอร์ ด้วยหน้าต่างบริบทโทเค็น 262K และการถอดรหัสการอนุมานแบบเร่ง ในเกณฑ์มาตรฐาน 15 รายการ ตัวแปรไตรภาคยังคงรักษาประสิทธิภาพพื้นฐานที่มีความแม่นยําเต็มที่ไว้ 95% ในขณะที่ตัวแปร 1 บิตยังคงรักษาไว้ 90% โดยแทบไม่มีการสูญเสียความสามารถทางคณิตศาสตร์และการเข้ารหัส โอเพ่นซอร์สภายใต้ใบอนุญาต Apache 2.0","category":"模型更新","source":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Bonsai 27B: โมเดลมัลติโมดอลระดับ 27B รุ่นแรกที่ทํางานบนสมาร์ทโฟน - ข่าว AI Aioga","description":"Bonsai 27B ใช้ Qwen 3.6 27B ซึ่งมีสองรุ่น: ไตรภาค (1.71 บิตที่ใช้งานจริง/น้ําหนัก, 5.9 GB) และ 1 บิต (1.125 บิตที่ใช้งานจริง/น้ําหนัก, 3.9 GB) ซึ่งเป็นรุ่นแรกที่นํารุ่นระดับ 27B มา","url":"https://www.aioga.com/th/news/cmrl2m4yg00awbicqo2arojvd/","contentTranslated":true,"sourceHash":"40b647793fcd518e","translatedAt":"2026-07-19T12:15:12.287Z"},"pl":{"title":"Bonsai 27B: Pierwszy multimodalny model 27B działający na smartfonie","summary":"Bonsai 27B opiera się na Qwen 3.6 27B, oferując dwa warianty: trójkątny (1,71 efektywnych bitów/masa, 5,9 GB) oraz 1-bitowy (1,125 efektywnych bitów/masa, 3,9 GB), z których ten drugi jako pierwszy wprowadza model 27B do iPhone'a 17 Pro. Model obsługuje wnioskowanie wieloetapowe, strukturalne wywołania narzędzi, zadania wizualne oraz pętle agentów do użytku komputerowego, oferując okno kontekstowe tokenów o pojemności 262K oraz przyspieszone dekodowanie wnioskowania. W 15 testach wariant trójwymiarowy zachował 95% pełnej precyzji bazowej, podczas gdy wariant 1-bitowy zachował 90%, praktycznie bez utraty możliwości matematycznych i kodowniczych. Open source na licencji Apache 2.0.","category":"模型更新","source":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Bonsai 27B: Pierwszy multimodalny model 27B działający na smartfonie - Aioga Wiadomości AI","description":"Bonsai 27B opiera się na Qwen 3.6 27B, oferując dwa warianty: trójkątny (1,71 efektywnych bitów/masa, 5,9 GB) oraz 1-bitowy (1,125 efektywnych bitów/masa, 3,9 GB), z których ten dr","url":"https://www.aioga.com/pl/news/cmrl2m4yg00awbicqo2arojvd/","contentTranslated":true,"sourceHash":"40b647793fcd518e","translatedAt":"2026-07-19T12:15:11.730Z"}}}}