Mira Murati's AI Startup Just Released Its First Model. It Wants Companies to Own Their AI, Not Rent It.

Key points

• Thinking Machines Lab released Inkling on Wednesday, its first publicly available AI model, roughly nine months after the company began building.
• Inkling has 975 billion total parameters (a measure of model complexity) but activates only about 41 billion for any single task, keeping it faster and cheaper to run than its size suggests.
• The model is open-weight, meaning any business or developer can download and modify it for free, unlike ChatGPT or Gemini.
• A joint study published in late June found that a custom model built using similar open-weight methods scored 84.7% on financial reasoning tests at roughly one-fourteenth the running cost of top proprietary AI models.
• Thinking Machines plans to make money through Tinker, its paid platform for training and adapting the model, not from the model download itself.

Mira Murati spent years as chief technology officer at OpenAI. Now she's betting against the thing OpenAI sells.

Thinking Machines Lab, the startup Murati founded after leaving OpenAI, released its first AI model on Wednesday. It's called Inkling, and it works differently from the headline AI products most people have heard of.

ChatGPT and Google Gemini are closed systems. You pay to use them. The company that built them keeps full control. Inkling is open-weight, which means any business or developer can download the full model and reshape it to fit their own needs, at no charge from Thinking Machines.

That's the whole pitch: AI that a business trains on its own data will outperform a general-purpose model a big lab sells to everyone.

The evidence backing it is genuinely interesting. Bridgewater Associates, the world's largest hedge fund, worked with researchers to take an existing open-source AI model and train it further on Bridgewater's own financial knowledge. That custom model scored 84.7% on financial reasoning tests, beating top proprietary AI products, at roughly one-fourteenth the running cost. The results come from a study the two companies published jointly in late June, not from an independent reviewer, so treat them with appropriate scepticism.

Microsoft chief executive Satya Nadella made a related point in a blog post on Sunday. Companies using closed AI products, he wrote, pay twice: once in subscription fees, and again by handing over business knowledge baked into their prompts and corrections, knowledge that can feed back into future model versions they don't own.

We covered the broader shift toward open-weight models just two days before this release, reporting that Chinese open-weight models took 41% of Hugging Face downloads this spring.

So is Inkling actually any good?

Thinking Machines says plainly that Inkling isn't the strongest AI model available, closed or open. It aims for steady, well-rounded performance at lower cost. The company says Inkling uses one-third as many processing steps as Nvidia's Nemotron 3 Ultra to reach the same score on coding tests. It can flag uncertainty rather than guessing, and users can dial its thinking effort up or down depending on whether they want a careful answer or a quick one.

Inkling is enormous under the hood, trained on 45 trillion tokens (the chunks of text, image and video data that AI systems learn from). Because of its mixture-of-experts architecture, where only a slice of the model activates per task, it stays practical to run.

Thinking Machines has been careful about one thing: Inkling was partly trained using outputs from other open-weight AI models, including Moonshot AI's Kimi K2.5, a common practice called distillation. The company says its next model will drop that approach entirely.

Where does the revenue actually come from?

Not from the model download. Revenue comes from Tinker, the paid platform for helping businesses fine-tune and host their own custom versions of Inkling. It's a sensible split: give away the weights, sell the expertise to put them to work.

Here's the honest judgement on this one. The Bridgewater numbers are compelling but self-reported, and Thinking Machines hasn't said how it balances its computing costs against revenue that isn't yet a primary focus. What's real is the direction: if your organisation has specific expertise locked inside past work and your team's heads, a custom model trained on that knowledge may genuinely outperform a general one at a fraction of the cost. The path to that is now a bit shorter.