Thinking Machines Lab released its first AI model Wednesday, called Inkling — a 975 billion parameter mixture-of-experts system that organizations can download and modify directly.

The startup, founded by former OpenAI CTO Mira Murati, designed Inkling as an open-weight alternative to closed models from major labs. Unlike flagship offerings from Anthropic or Google, developers can access and adapt the model's weights for specific use cases.

Inkling uses only 41 billion of its total parameters for any given task, a design choice that keeps the large model faster and cheaper to run. The company trained it on 45 trillion tokens spanning text, image, audio, and video data, though current outputs are limited to text, code, and structured data.

The model represents Thinking Machines' first public proof point after 18 months building AI infrastructure largely out of view. The company previously showcased "interaction models" in May — AI systems designed to listen, speak, and interrupt rather than follow traditional chatbot patterns.

Testing the customization thesis

Inkling embodies Thinking Machines' central bet that customizable AI will outperform one-size-fits-all models from major labs. The model includes calibrated uncertainty — flagging when it's unsure rather than guessing — and allows users to adjust "thinking effort" to trade accuracy for speed.

On coding benchmarks, Thinking Machines claims Inkling uses one-third the tokens of Nvidia's Nemotron 3 Ultra to achieve equivalent performance. The company acknowledges Inkling isn't "the strongest overall model available today, open or closed," but positions it as a foundation for specialized applications.

The release tests whether enterprises will choose adaptable models over the general-purpose systems currently dominating the market. Thinking Machines argues that organizations need AI they can modify for specific domains rather than relying on broad-purpose alternatives.

Inkling is available for download through the company's website, with commercial licensing terms allowing modification and redistribution. The startup plans additional model releases as it builds toward more specialized AI systems for enterprise customers.