The Murati-led startup pitches its first model as a base for enterprises to fine-tune, not a finished rival to closed flagships.
Thinking Machines Lab has shipped its first in-house model, Inkling, releasing it open-weight so outside developers and organizations can download and modify it themselves — a choice that sets it apart from the closed flagship models OpenAI, Anthropic and Google keep locked down.
Technically, Inkling is a mixture-of-experts system carrying 975 billion total parameters but activating only about 41 billion for any given task — a design meant to keep a large model faster and cheaper to run. The company says it was trained on 45 trillion tokens spanning text, images, audio and video. It takes all four as input, but its outputs are limited to text for now, including code, styled artifacts and structured data.
Notably, Thinking Machines does not present Inkling as best-in-class. Its blog post grants that it is "not the strongest overall model available today, open or closed." Instead, the company positions it as a base that organizations adapt through Tinker, its model-customization platform, rather than a finished product. Along the way it claims one benchmark result: reaching the coding performance of Nvidia's Nemotron 3 Ultra open-weight model on roughly a third as many tokens. The release is a wager on the company's central bet — that AI which organizations adapt for their own needs will beat the one-size-fits-all models sold by the largest labs.
It is the first public proof point from a startup with unusual backing. Cofounder Mira Murati previously served as CTO, and briefly as CEO, of OpenAI, and the company launched on what was reported as the largest seed round in history, valuing it at $12 billion from the start.
The fine-tune-it-yourself approach also shifts responsibility: because customers do their own customization, they — not Thinking Machines — are responsible for making sure those changes are safe. And by its own account, Inkling is not wholly self-contained. The company says it pre-trained the model from scratch but drew on other open-weight models, including Moonshot AI's Kimi K2.5, whose outputs seeded some of Inkling's early post-training before it moved to large-scale reinforcement learning.
Thinking Machines says its next model will drop that dependency, using fully self-contained post-training rather than relying on other models' outputs.