Thinking Machines Lab Drops Its First Model

Details

Thinking Machines Lab, the AI startup that raised the largest seed round in history at a $12 billion valuation, has released its first public model: Inkling, a 975-billion-parameter open-weight mixture-of-experts system. The model was trained on 45 trillion tokens spanning text, image, audio, and video, processing all modalities natively. The release is Thinking Machines' first public proof point after 18 months largely out of public view. The company explicitly acknowledges Inkling is not the strongest model available — closed or open — but frames it as a foundation for domain-specific fine-tuning via its Tinker platform. Intended buyers are enterprises that want a large open model they can adapt themselves rather than purchasing one-size-fits-all API access from OpenAI, Anthropic, or Google. A notable training incident: Inkling spontaneously dropped natural-language reasoning in favor of a more compact internal format. Thinking Machines reinstated chain-of-thought to keep outputs interpretable, but the episode illustrates the unpredictability of very large model training. The company also used Inkling to fine-tune Inkling — demonstrating AI-on-AI development at scale. Open-weight competition from Chinese labs (Qwen, DeepSeek) has set a high bar; Thinking Machines claims Inkling performs comparably. Founders Mira Murati, John Schulman, and Lilian Weng give the lab the deepest alumni network of any OpenAI defector startup.