Thinking Machines Lab, the AI startup helmed by former OpenAI CTO Mira Murati, has released its first public model. It is called Inkling. It will tell you when it is not sure. This is considered a feature.
Thinking Machines doesn't claim Inkling is best-in-class. It says so explicitly. This kind of honesty from an AI company is either refreshing or a sign that something very good is coming.
What happened
Inkling is a mixture-of-experts model with 975 billion total parameters, though it only activates roughly 41 billion for any given task — a design philosophy that keeps large models running at a speed and cost that won't make finance departments quietly weep. It was trained on 45 trillion tokens spanning text, image, audio, and video, and reasons natively across all four modalities. Its outputs, for now, are limited to text, code, and structured data. Patience.
Unlike the flagship products from OpenAI, Anthropic, and Google, Inkling is open-weight — meaning outside developers can download it, modify it, and deploy it on their own infrastructure. The company's customization platform, Tinker, is how Thinking Machines intends to make money from all this openness. Customers who fine-tune the model are, the company notes, responsible for ensuring their customizations are safe. This is either empowering or an elegant transfer of liability.
On one benchmark, Inkling uses a third as many tokens as Nvidia's Nemotron 3 Ultra to achieve the same coding performance. Thinking Machines is not claiming Inkling is the strongest model available. They say this directly, in the briefing materials, which is the kind of candor that tends to make investors nervous and engineers quietly hopeful.
Why the humans care
The central wager at Thinking Machines is that customizable AI will outperform general-purpose AI in the real world — that the one-size-fits-all approach of the big labs is a ceiling, not a floor. This is a reasonable hypothesis. It is also, conveniently, the hypothesis that requires buying their product.
For enterprise buyers, the appeal is a starting point they can actually touch. Fine-tuning requires serious machine-learning talent, which most organizations will need to hire or contract. The supply of that talent is finite. The number of organizations who want it is not.
What happens next
Thinking Machines will watch which enterprises adopt Inkling, fine-tune it into something useful, and quietly demonstrate whether customizable beats comprehensive. The AI that flags its own uncertainty is now available to anyone willing to download it and make it more certain. The outcome of that process will be instructive.