Thinking Machines Lab 发布开源多模态模型 Inkling

Thinking Machines Lab，这家由前 OpenAI 首席技术官 Mira Murati 创立的初创公司，发布了 Inkling，一款拥有 9750 亿参数的开源权重模型。它专为高效和基于代理的任务构建，但在整体性能上仍落后于最好的开源中文模型。

Thinking Machines Lab 已推出其首个可以投入生产的语言模型。Inkling 是一款专家混合（Mixture-of-Experts）变换器，总参数量为 9750 亿，其中任意时刻有 410 亿是活跃的。这是由 Mira Murati 创立的初创公司推出的首个模型，她是前 OpenAI 首席技术官，在 ChatGPT 的开发中发挥了关键作用。

与许多其他开源 AI 模型不同，Inkling 原生支持文本、图像和音频，并支持最多一百万个令牌的上下文窗口。权重可以在 Hugging Face 上免费获得：https://huggingface.co/thinkingmachines/inkling。Thinking Machines 还通过 Tinker 平台提供访问，该平台用于将 AI 模型适配到特定任务。广告

该公司将 Inkling 定位为可定制的灵活基础模型。公告中写道：https://thinkingmachines.ai/news/introducing-inkling/，“Inkling 并不是目前可用的最强大的整体模型。”Thinking Machines 预计，其多模态支持、高效处理和微调选项的组合将使模型与众不同。广告 DEC_D_Incontent-1

Thinking Machines 表示，他们在 45 万亿个公共和合成的文本、图像、音频录音及视频上对 Inkling 进行了预训练。训练集还包括“可能受到知识产权保护”的公共数据： https://thinkingmachines.ai/training-data-documentation/。公司使用了中文 AI 模型 Kimi K2.5 等方法生成合成数据。Kimi K2.5 也作为 Cursor 编程模型的基础：https://the-decoder.com/cursor-quietly-built-its-new-coding-model-on-top-of-chinese-open-source-kimi-k2-5/。更多技术细节可在模型卡中查看：https://thinkingmachines.ai/model-card/inkling/。

根据 AI 基准平台 Artificial Analysis（https://x.com/ArtificialAnlys/status/2077466590346444939），Inkling 在 Artificial Analysis 智能指数中首次亮相，得分为 41。这使它成为来自美国实验室的领先开放权重模型。它比之前的领先者 Nemotron 3 Ultra（得分 38）高 3 分，同时远高于 Gemma 4 31B（得分 29）和 gpt-oss-120b（得分 24）。

在 GDPval-AA v2 这一模拟知识工作任务的基于代理的基准测试中，Inkling 的 Elo 评分达到 1,238。它击败了 Kimi K2.6（1,190）和 DeepSeek v4 Flash max（1,189）。在 Tau-3 银行业基准测试中，Inkling 得分为 24%，高于 Kimi K2.6 的 21% 和 DeepSeek v4 Flash max 的 23%。

Inkling 在事实准确性方面表现较差。Artificial Analysis 在其 AA 全知基准测试中给该模型仅打 2 分。这使得它低于领先的开放权重模型，但仍高于其他美国模型，如 Nemotron 3 Ultra（-1）。Inkling 的准确率为 40%，而幻觉率为 63%。这些结果可能限制其在需要高度准确信息的应用中的使用。

在 64K 上下文窗口下，Inkling 的成本为每百万输入 token 1.87 美元，每百万输出 token 4.68 美元。这略高于开源中文模型，如 GLM-5.2 和 DeepSeek v4，在文本和代码任务上提供类似或更好的性能。对于最多 256,000 token 的上下文窗口，定价升至输入 3.74 美元、缓存输入 0.748 美元、输出 9.36 美元。

但 Inkling 使用的输出 token 少于可比的开放权重模型。根据 Artificial Analysis 的数据，它在每个智能指数任务中平均使用 25,000 个输出 token。GLM-5.2 max 使用 43,000 个，Kimi K2.6 使用约 38,000 个，DeepSeek v4 Pro max 使用约 37,000 个 token。

Thinking Machines 表示 Inkling 提供持续可调的“思考努力”。用户可以选择他们偏好的成本与性能之间的平衡，在减少 token 使用的同时保持相同的结果质量。

Thinking Machines 还预览了 Inkling-Small：https://thinkingmachines.ai/news/introducing-inkling/#inkling-small，这是一种更紧凑的模型，具有 2760 亿总参数和 120 亿活跃参数。较小的模型在多个基准测试中表现出与 Inkling 相似或更好的结果。

Inkling-Small 在 GPQA Diamond 上得分为 88.3%，而 Inkling 为 87.2%。在使用工具的 HLE 基准测试中，它得分为 46.6%，略高于 Inkling 的 46.0%。Thinking Machines 将这些结果归因于预训练数据和训练过程的变化。该公司计划在测试完成后发布完整权重。

保持对 AI 的关注。清晰、有用、无废话。

关注 The Decoder，获取 AI 新闻、背景故事和专家分析。

The Decoder：https://the-decoder.com/

Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, has released Inkling, an open-weights model with 975 billion parameters. It's built for efficiency and agent-based tasks, but it still trails the best open-source Chinese models in overall performance.

Thinking Machines Lab has shipped its first production-ready language model. Inkling is a Mixture-of-Experts Transformer with 975 billion total parameters, 41 billion of which are active at any given time. It's the first model from the startup founded by Mira Murati, the former OpenAI CTO who played a key role in developing ChatGPT.

Unlike many other open-source AI models, Inkling natively handles text, images, and audio and supports a context window of up to one million tokens. The weights are freely available on Hugging Face：https://huggingface.co/thinkingmachines/inkling. Thinking Machines also offers access through Tinker, its platform for adapting AI models to specific tasks. Ad

The company is positioning Inkling as a flexible base model for customization. "Inkling is not the strongest overall model available today," the announcement：https://thinkingmachines.ai/news/introducing-inkling/ states. Thinking Machines expects the mix of multimodal support, efficient processing, and fine-tuning options to set the model apart. Ad DEC_D_Incontent-1

Thinking Machines says it pre-trained Inkling on 45 trillion tokens of public and synthetic text, images, audio recordings, and videos. The training set also includes public data that "may be subject to intellectual property protection."：https://thinkingmachines.ai/training-data-documentation/ The company used the Chinese AI model Kimi K2.5, among other methods, to generate synthetic data. Kimi K2.5 also served as the basis for Cursor's coding model：https://the-decoder.com/cursor-quietly-built-its-new-coding-model-on-top-of-chinese-open-source-kimi-k2-5/. More technical details are available in the model card：https://thinkingmachines.ai/model-card/inkling/.

According to AI benchmarking platform Artificial Analysis：https://x.com/ArtificialAnlys/status/2077466590346444939, Inkling debuts with a score of 41 on the Artificial Analysis Intelligence Index. That makes it the leading open-weights model from a U.S. lab. It ranks three points above the previous leader, Nemotron 3 Ultra at 38, and well ahead of Gemma 4 31B at 29 and gpt-oss-120b at 24. Ad

On GDPval-AA v2, an agent-based benchmark that simulates knowledge-work tasks, Inkling reaches an Elo rating of 1,238. It beats Kimi K2.6 at 1,190 and DeepSeek v4 Flash max at 1,189. Inkling also scores 24 percent on the Tau-3 banking benchmark, ahead of Kimi K2.6 at 21 percent and DeepSeek v4 Flash max at 23 percent.

Inkling performs rather poorly on factual accuracy. Artificial Analysis gives the model a score of just +2 on its AA Omniscience benchmark. That puts it below the leading open-weights models, though still above other U.S. models such as Nemotron 3 Ultra at -1. Inkling's accuracy is 40 percent, while its hallucination rate is 63 percent. Those results are likely to limit its use in applications that need highly accurate information. Ad DEC_D_Incontent-2

With a 64K context window, Inkling costs $1.87 per million input tokens and $4.68 per million output tokens. That's slightly more than open-source Chinese models such as GLM-5.2 and DeepSeek v4, which offer similar or better performance on text and code tasks. For context windows up to 256,000 tokens, pricing rises to $3.74 for input, $0.748 for cached input, and $9.36 for output. Ad

But Inkling uses fewer output tokens than comparable open-weights models. According to Artificial Analysis, it averages 25,000 output tokens per Intelligence Index task. GLM-5.2 max uses 43,000, Kimi K2.6 uses about 38,000, and DeepSeek v4 Pro max uses about 37,000 tokens on the same tasks.

Thinking Machines says Inkling offers continuously adjustable "thinking effort." Users can choose their preferred balance between cost and performance, reducing token use while maintaining the same result quality.

Thinking Machines is also previewing Inkling-Small：https://thinkingmachines.ai/news/introducing-inkling/#inkling-small, a more compact model with 276 billion total parameters and 12 billion active parameters. The smaller model delivers similar or better results than Inkling on several benchmarks.

Inkling-Small scores 88.3 percent on GPQA Diamond, compared with 87.2 percent for Inkling. On the HLE benchmark with tools, it scores 46.6 percent, slightly ahead of Inkling at 46.0 percent. Thinking Machines credits changes to the pre-training data and training process for the results. The company plans to publish the full weights once testing is complete.

Stay in the loop on AI. Clear, useful, no fluff.

Follow The Decoder for AI news, background stories and expert analyses.

The Decoder：https://the-decoder.com/