Thinking Machine Lab, founded by former OpenAI Chief Technology Officer Miriam Murati, has released its first multi-modal AI model, Inkling, trained from scratch. It is regarded as the strongest open-source model in the United States currently. The model uses a mixture-of-experts architecture, with a total of 975B parameters and 41B activated parameters. It supports up to 1 million tokens of context length and was pre-trained on 45 trillion tokens of data, covering four modalities: text, images, audio, and video. The model's weights are now available for download.

image.png

Team is formidable, but reasoning and programming still lag behind Chinese open-source models

About two-thirds of the core members of Thinking Machine Lab come from OpenAI. The team previously led cutting-edge research, products, and security work at the organization, making their lineup extremely impressive. However, despite this, Inkling still lags behind Chinese open-source models in reasoning and programming. GLM 5.2 leads comprehensively in coding, agent reasoning, and complex reasoning tasks, with a SWEBench Pro score of 62.1% versus 54.3%, and Terminal Bench 2.1 scores 82.7% versus 63.8%.

DeepSeek V4 Pro maintains an advantage in coding and factual tests, and Kimi K2.6 also wins in multiple technical benchmark tests. However, Inkling surpasses DeepSeek in mathematics, achieving an impressive score of 97.1% on the AIME2026 test.

Outstanding multimodal capabilities, but still lags behind closed-source giants

Compared to local open-source competitors in the United States, Inkling shows significant advantages, fully dominating NVIDIA's Nemotron 3 Ultra in reasoning, coding, and agent workflows. However, when facing closed-source giants like Claude Fable 5, GPT-5.6 Sol, and Gemini 3.1 Pro, Inkling still has a clear gap in extreme reasoning and software engineering autonomy. Claude Fable 5 achieved 95.0% on SWEBench Verified, and GPT-5.6 Sol scored 89.5 on Terminal Bench 2.1, both far exceeding Inkling.