What Is Thinking Machines Lab? Inside Mira Murati's AI Startup and Its Bet on Customization

When Mira Murati left OpenAI in September 2024 after six and a half years as its chief technology officer, the obvious question was what she would do next. The answer, unveiled in February 2025, was Thinking Machines Lab: a new AI company that raised one of the largest seed rounds in venture history before it had shipped anything, and that in July 2026 released its first model, an open-weights system called Inkling, roughly nine months after starting to build in earnest.

Thinking Machines is worth understanding not because it is the biggest lab, it is not, but because it represents a specific and contrarian bet about where AI value will accrue. This piece profiles the company neutrally: who is behind it, how it is funded and governed, what it has actually shipped, and what its strategy reveals about a fault line running through the whole industry.

$2BSeed round, July 2025at a $12B valuation, led by Andreessen Horowitz

Feb 2025Foundedby ex-OpenAI CTO Mira Murati

~9 monthsFounding to first shipped modelInkling, July 2026

The people

The founding story is, in large part, an OpenAI alumni story. Murati assembled a founding group heavy with former OpenAI researchers and leaders, including John Schulman, an OpenAI co-founder who became chief scientist at Thinking Machines after a brief spell at Anthropic, along with Lilian Weng, Barret Zoph, Andrew Tulloch and Luke Metz. The early team also drew talent from Meta, Mistral and elsewhere, and the company added advisers including former OpenAI research leaders.

That pedigree is the company's clearest asset and its clearest vulnerability. The asset is obvious: this is a concentration of people who have built frontier systems before. The vulnerability showed in January 2026, when two of the founding group departed back to OpenAI and another left for Meta's superintelligence effort. In a field where a handful of researchers can move the frontier, and where rivals preparing to go public can offer compelling equity, talent is both the moat and the thing most easily poached. The company has publicly emphasised continuity over dependence on any single personality, which reads as a direct response to exactly this risk.

The money and the governance

The financial story is unusual even by 2025 standards. In July 2025 Thinking Machines closed a 2 billion dollar seed round at a 12 billion dollar valuation, led by Andreessen Horowitz, with Nvidia, AMD, Cisco, Jane Street and others participating. The round was widely reported as one of the largest seeds ever, and it even included a 10 million dollar investment from the government of Albania, Murati's country of origin. By late 2025, reporting placed the company in conversations at a valuation as high as 50 to 60 billion dollars, though a reported larger round was said to have stalled around the start of 2026, and the company has declined to discuss its current funding status.

Thinking Machines Lab, from founding to first model

Key milestones on the path from a February 2025 launch to shipping an open-weights foundation model.

Feb 2025Founded
Mira Murati, former OpenAI CTO, founds Thinking Machines Lab in San Francisco as a public benefit corporation, with a founding team drawn largely from OpenAI.
Jul 2025$2B seed at $12B
Closes a $2 billion round led by Andreessen Horowitz, reported as one of the largest seed rounds ever, with Nvidia, AMD, Cisco and Jane Street among backers.
Oct 2025Tinker launches
Ships Tinker, an API for fine-tuning open language models. The customization platform, not model licensing, is the intended revenue engine.
Jan 2026Departures
Two figures from the founding group depart back to OpenAI, and another leaves for Meta, testing the company’s stated emphasis on continuity over any single personality.
Mar 2026Nvidia compute deal
Announces a strategic Nvidia partnership to deploy around a gigawatt of computing capacity, the infrastructure on which Inkling was later trained.
Jul 2026Inkling ships
Releases Inkling, its first model, as open weights under the Apache license, roughly nine months from founding to a shipped foundation model.

Sources: Wikipedia, TechCrunch, Bloomberg and company announcements. Valuation and funding figures are as reported.

The governance is as notable as the numbers. Thinking Machines is structured as a public benefit corporation, and Murati holds weighted voting control, with founding shares reportedly carrying votes weighted far more heavily than ordinary shares and a deciding vote on board matters. This concentrates strategic control in the founder to an unusual degree. There is a reasonable case for it, protecting a long-horizon research mission from short-term investor pressure, and a reasonable concern about it, that heavy founder control reduces external accountability. Both readings are legitimate, and which one matters more will only be clear in how the control is used over time.

What it has actually shipped

A company can raise billions on reputation, but the more telling record is what reaches users. Thinking Machines has shipped two things that matter.

The first, in October 2025, was Tinker, an API for fine-tuning open language models. Tinker is easy to underrate as a developer tool, but it is central to the company's entire business logic. The second, in July 2026, was Inkling, its first foundation model, released as open weights under the Apache license. Notably, the company reached a shipped foundation model in roughly nine months, faster than the multi-year runways OpenAI and Anthropic took to their first market proof, a pace worth acknowledging even while remembering that shipping a capable model and building a durable business are different achievements.

The strategy only makes sense as a pair. Give the model away, and sell the platform that makes the model yours.

The thesis: customization over one-size-fits-all

The reason Tinker and Inkling belong together is the company's central bet. Most frontier labs sell access to a single, centrally trained, frozen model that every customer rents through an API. Thinking Machines argues that this is the wrong shape for a lot of enterprise value, because, in the company's framing, much valuable expertise is specific to the people and organisations that hold it, and a one-size-fits-all model cannot absorb it. Its answer is to release capable open weights and sell the infrastructure, Tinker, that lets organisations fine-tune and own a customised version.

This is a genuinely different revenue model. OpenAI and Anthropic monetise primarily through per-token usage of proprietary models. Thinking Machines makes the weights free and aims to earn from the training, fine-tuning and hosting ecosystem around them. The strategic logic echoes a broader argument, voiced even by executives at large proprietary vendors, that organisations relying wholly on closed models can end up paying twice, once in fees and again by exporting their proprietary knowledge into someone else's system.

It is important to weigh this neutrally rather than adopt it. The customization thesis has real force: specialised, fine-tuned models can outperform general ones on narrow tasks at lower cost, and one reported partnership with an investment firm produced a fine-tuned open model that beat proprietary competitors on a financial-reasoning test at a fraction of the operating cost. But the thesis also has real limits: many organisations lack the skill or appetite to fine-tune and operate their own models, the very general-purpose convenience Thinking Machines is betting against is exactly why the big chatbots reached hundreds of millions of users, and open weights shift the burden of safety and correctness onto the customer. The honest position is that this is a plausible, unproven bet on one segment of the market, not a settled verdict about where all value will land.

Where Thinking Machines sits in the landscape

It helps to place the company precisely. It is not trying to win the raw-capability crown; it has said outright that its first model is not the strongest available. It is not a consumer chatbot company chasing OpenAI's distribution. It is positioning as the customization layer for organisations that want to adapt strong models to their own data and workflows, and keep them, rather than rent a frozen model from a frontier lab.

That places it on one side of the defining tension of 2026: closed, centrally controlled frontier models on one hand, and open, adaptable, portable models on the other. Thinking Machines is a well-capitalised, credibly staffed bet on the second. The wider ecosystem is arranging itself around the same question, and the tooling that lets teams actually exploit both worlds is becoming part of the standard stack. A model-agnostic workspace such as Metir AI, which brings open-weights and closed models together in one place, exists precisely because the answer for most teams is not one camp or the other but the freedom to use whichever fits a given task. Thinking Machines is a bet that a large share of that demand tilts toward the adaptable, ownable side.

The bigger picture

Thinking Machines Lab is one of the most watched companies in AI for reasons that have less to do with its size than with its stance. It pairs an unusually strong founding team with an unusually large seed, an unusual governance structure, and an unusually contrarian thesis, then backs all of it by actually shipping. Whether the bet pays off depends on questions that are still open: whether enough organisations want to own and customise models rather than rent them, whether Tinker can become a durable business around free weights, and whether the company can retain the people who make it credible.

None of those questions is answered yet, which is the accurate place to leave it. What can be said is that Thinking Machines has made the customization-versus-centralisation debate concrete, put serious money and talent behind one side of it, and given the rest of the industry something real to measure against.

Use the adaptable and the frontier models, in one place

Thinking Machines is betting that teams want models they can shape and own, not just rent. The practical way to act on that is to keep every option open. Metir AI brings open-weights and closed frontier models into a single workspace, so you can fine-tune and own where it pays and reach for a frontier model where it does not, without rebuilding your stack. Try Metir AI free and stay on the whole field.

Sources: