What Is Thinking Machines Lab?

Thinking Machines Lab is an AI startup led by CEO Mira Murati, the former technology chief at OpenAI. It builds customizable, open-weights AI models and tools that let developers train and run those models without owning a supercomputer.

The lab wants to loosen the grip that closed frontier labs hold over AI. It argues that AI should be decentralized and built on local knowledge, not planned centrally by a few large companies.

This guide explains what Thinking Machines Lab is, who runs it, and what it builds. It also covers its Nvidia partnership and whether its models make sense for a regulated small business.

What Is Thinking Machines Lab?

Thinking Machines Lab is an AI startup that builds open-weights models and developer tools. It is led by CEO Mira Murati, who was OpenAI's chief technology officer before founding the company.

Open-weights means the model's trained parameters are published for others to download and run. That is different from closed models like GPT-5 or Claude, which you can only reach through a vendor's API.

The lab has two main outputs so far. Tinker is a cloud fine-tuning tool. Inkling is its first AI model, released on July 15, 2026.

On Friday the lab published its first manifesto. The document outlines a future where AI is decentralized and built on local knowledge rather than controlled by a handful of frontier labs.

Who Is Mira Murati?

Mira Murati is the CEO and founder of Thinking Machines Lab. She served as chief technology officer at OpenAI, where she helped lead the work behind ChatGPT.

Murati grew up in Albania and witnessed the collapse of communism there as a child. She has said that experience shapes her view of how power should be distributed.

She frames closed AI labs as a form of central planning. In her view, a few companies deciding how AI works for everyone repeats a pattern she saw fail up close.

That belief drives the lab's mission. Murati wants AI to be built locally and shaped by the people who use it, not dictated from the top down.

The Mission: Decentralized, Open-Weights AI

Thinking Machines Lab's mission is to loosen the grip that frontier labs hold over AI. It does this by releasing more customizable, open-weights models that anyone can download and adapt.

Closed labs like OpenAI and Anthropic keep their model weights private. You rent access through their systems, and you cannot inspect or fully control the model.

Open-weights flips that arrangement. A hospital, bank, or law firm can host the model itself, tune it on private data, and keep sensitive records inside its own walls.

For regulated businesses, that control matters. Data residency, audit trails, and vendor lock-in all get easier to manage when you own the model you run.

Its Products: Tinker and Inkling

Thinking Machines Lab has shipped two products: the Tinker fine-tuning tool and the Inkling model. Together they let a developer customize a large model without managing heavy infrastructure.

Tinker came first, released last year. Inkling is the newer piece, and it is the lab's first model of its own.

• Tinker — a cloud-based fine-tuning tool and API. It lets a developer at a laptop customize and train large models without running a supercomputing cluster. The lab manages the hard infrastructure so your team can focus on the data.
• Inkling — the lab's first AI model, released on Wednesday, July 15, 2026. It is an open-weights mixture-of-experts (MoE) model with 975 billion total parameters and 41 billion active per token, so it draws on a huge base while keeping each request efficient.

Nvidia Partnership and Backing

Thinking Machines Lab is backed by Nvidia, which invested in the startup in a partnership announced in March. The two also struck a large hardware deal.

Under that deal, the lab agreed to deploy at least one gigawatt of cutting-edge chips to train and serve its frontier models. That is a very large amount of computing power.

The lab trained Inkling entirely on Nvidia hardware. The chip partnership is what makes a model of that size possible for a young company.

Funding is context here, not the headline. What matters for buyers is that the lab has the compute and backing to keep shipping and supporting its models.

Should Your Business Use Thinking Machines' Models?

Thinking Machines' models fit businesses that need control over their data and their model. The open-weights approach is strongest when privacy, cost, or customization are the priority.

There is a real proof point. Hedge fund Bridgewater Associates used Tinker to fine-tune the open-weights Qwen3-235B model on its own data.

Bridgewater said the tuned model beat GPT-5 and Claude Opus on financial-document triage. It also cut compute costs by more than 13 times.

Still, open-weights is not a fit for every team. You need people who can host, tune, and monitor a model, or a partner who can do it for you.

• Good fit if you handle sensitive records and want the model inside your own environment.
• Good fit if you have a narrow, repeatable task where a tuned smaller model can beat a general one.
• Weaker fit if you have no engineering support and just need a chatbot out of the box.
• Weigh safety too: the lab tested Inkling for bio and cyber misuse risks and is still studying open-weights safeguards.

How Thinking Machines Differs From OpenAI and Anthropic

Thinking Machines Lab differs from OpenAI and Anthropic mainly on openness. It publishes model weights, while those labs keep theirs closed.

With OpenAI or Anthropic, you send your data to their servers and get an answer back. You do not hold the model, and you cannot move it into your own systems.

With Thinking Machines Lab, you can download the weights and run the model where you choose. That trades some convenience for far more control.

The philosophy is different too. Where the big labs concentrate power, Thinking Machines Lab argues for spreading it out. For a regulated buyer weighing risk, that difference in control can outweigh raw benchmark scores.