What launched
Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, released its first proprietary model on July 15: Inkling, a mixture-of-experts system with 975 billion total parameters, of which roughly 41 billion are active for any given task. It was trained on 45 trillion tokens spanning text, image, and audio, reasons natively across those inputs, and supports a context window up to 1 million tokens. Unlike the flagship chatbots from OpenAI, Anthropic, and Google, Inkling ships open-weight under an Apache 2.0 license, so outside developers and companies can download and modify it directly.
A standout feature is “controllable thinking effort” — a dial that lets users trade reasoning depth for speed and token cost, alongside calibrated answers that are designed to flag uncertainty rather than guess. On one benchmark the company cites, Inkling matches Nvidia’s Nemotron 3 Ultra on coding performance using roughly a third of the tokens. Thinking Machines is candid that Inkling isn’t best-in-class on raw capability — its own materials call it “not the strongest model available today, closed or open” — and is positioning it instead as a well-rounded starting point.
Why it matters
Inkling is the company’s first public proof point after a year and a half spent building AI infrastructure largely out of public view, and it’s a direct test of Thinking Machines’ central bet: that AI organizations adapt for themselves will outperform the one-size-fits-all models sold by the biggest labs. The company is marketing Inkling less as a finished product than as a base for enterprises to fine-tune through Tinker, its model-customization platform — meaning revenue is expected to come from training and fine-tuning services and the hosting ecosystem around the model, not metered access to the model itself.
That thesis got an early data point from a joint project with Bridgewater Associates: researchers from both companies fine-tuned an existing open-source model on Bridgewater’s own financial expertise and scored 84.7% on financial reasoning tests — beating top proprietary models while costing roughly a fourteenth as much to run, per the companies’ own (not independently verified) evaluation. For founders and technical teams deciding whether to build on a rented frontier API or invest in owning and customizing an open-weight model, Inkling adds a credible, well-funded option to the “self-host and fine-tune” side of that decision — particularly relevant as Microsoft CEO Satya Nadella and Hugging Face CEO Clem Delangue have both recently argued that enterprise AI workloads are shifting toward private and open-source models for exactly this reason.
Where to try it
Inkling’s weights and technical report are available via Thinking Machines Lab’s announcement post; fine-tuning and hosting are accessed through the company’s Tinker platform.