Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, has released its first proprietary AI model, Inkling, as an open-weight system that outside developers and companies can download and modify. The launch matters because it is an early public test of the company’s core argument that organisations will get better results from adapting AI to their own needs rather than relying on standardised models sold by the biggest labs. Rather than claiming top performance overall, Thinking Machines is positioning Inkling as a flexible, enterprise-focused base model that can be customised and tuned.

Inkling is a mixture-of-experts model with 975 billion total parameters, though it uses about 41 billion for any given task, and the company says it was trained on 45 trillion tokens spanning text, images, audio and video. Thinking Machines says the model can express uncertainty instead of guessing and lets users adjust how much “thinking effort” it uses to balance speed and performance; on one benchmark, it reportedly matched Nvidia’s Nemotron 3 Ultra in coding while using a third as many tokens. The company plans for customers to refine it through its Tinker customisation platform, reflecting a wider industry debate over whether businesses should keep paying for centralised proprietary AI or build more value from models they control themselves.