Thinking Machines Lab is Playing a Different Game From Other Frontier AI Labs
Thinking Machines Lab just did something almost no frontier AI company does. It shipped a model and led with a disclaimer. Inkling, the company states plainly, is “not the strongest overall model available today, open or closed.” Instead, the company argues, a combination of qualities makes it a good open-weights base for customization. In an industry where every release is the smartest, fastest, most capable thing ever built, that sentence is the tell. Either this is a company that doesn’t know how to market, or it’s a company playing a different game entirely. Eighteen months of evidence points to the second reading, and the game they’re playing is the interesting part. (Thinking Machines Lab, TechCrunch)
The Counterintuitive Bet
Every major lab says roughly the same thing: build the most capable general-purpose model, sell access through an API, layer agents on top. OpenAI, Anthropic, and Google all built their flagship products as general-purpose chatbots first, with agentic and autonomous features added over time. That is the dominant axis, and everyone is optimizing along it. (TechCrunch)
Thinking Machines rotated ninety degrees. The central bet behind the startup is that AI an organization can adapt for itself will outperform the one-size-fits-all models the biggest labs currently sell. An essay the company published last week was clearly meant as the backdrop for this release. It argues that AI trained centrally by one company and then set in stone will underperform AI that organizations shape themselves, because the knowledge that matters is local and dispersed. The thesis is about where intelligence should live: not in one centrally aligned oracle, but in the hands of the people who actually hold the domain knowledge. (TechCrunch, Thinking Machines Lab)
From the start, the company has also distinguished itself by focusing on multimodal systems that work with people collaboratively rather than aiming for fully autonomous agents. While the rest of the industry races toward AI that works instead of you, Thinking Machines keeps building AI that works with you. That distinction determines what they build, as the product history shows. (Crunchbase News, Data Center Dynamics)
The Unique Company
The founding facts are by now well known. Thinking Machines Lab was founded in February 2025 by Mira Murati, the former chief technology officer of OpenAI. By July it had officially closed a $2 billion seed round led by Andreessen Horowitz, with participation from Nvidia, Accel, ServiceNow, Cisco, AMD, and Jane Street, valuing the startup at $12 billion. It was by far the largest seed round on record, closed before the company had revealed what it was working on. The company operates as a public benefit corporation, and by launch had assembled roughly 30 researchers and engineers from OpenAI, Meta, and Mistral AI, a founding group that included former OpenAI research VP Barret Zoph and OpenAI cofounder John Schulman. (TechCrunch, Crunchbase News)
One structural detail is worth noting. During the company’s fundraising, multiple reports described unusually strong control rights demanded by, and granted to, Murati. Combined with the public benefit structure, the picture is of a company built to resist capture, whether by a board coup, an acquirer, or investors demanding a pivot. That can be read as mission protection or as accountability-light governance. It is honestly both. (Spyglass)
The Challenging Year
An honest account has to include the turbulence, because there was plenty. Co-founder Andrew Tulloch departed for Meta in October 2025, during a stretch when Meta was reportedly offering compensation packages that climbed into the hundreds of millions of dollars. Then in January came the harder blow. Murati announced that the company had “parted ways” with co-founder and CTO Barret Zoph, and less than an hour later OpenAI’s CEO of applications announced that Zoph, co-founder Luke Metz, and researcher Sam Schoenholz were all returning to OpenAI, with additional researchers reported to be following. Around the same time, Bloomberg reported the company had been in discussions to raise new funding at a $50 billion valuation, a round that did not close, and reporting described internal frustration about a lack of clarity on product direction. (TechCrunch, Fortune)
Some observers wrote the company’s obituary in January. What matters is what happened instead. The company named Soumith Chintala, best known as a co-creator of PyTorch, as its new CTO. In March it announced a multi-year strategic partnership with Nvidia to deploy at least one gigawatt of next-generation Vera Rubin systems for frontier model training, with Nvidia also making a significant investment in the company. Then it shipped the model. The apparent resolution of the “tools versus models” argument turned out to be both, delivered as one product. That is the part the obituaries missed. (TechCrunch, Nvidia)
The Products: One Idea, Spoken Three Times
Viewed through the thesis, the product sequence stops looking scattered and starts looking deliberate.
Tinker came first. Announced October 1, 2025, Tinker is a flexible API for fine-tuning language models. Users write training loops in Python and the company runs them on its distributed GPU infrastructure, using Low-Rank Adaptation to cut the compute required to a fraction of full fine-tuning. The design won respect from serious people. Former OpenAI co-founder Andrej Karpathy praised its tradeoffs, writing that Tinker lets users retain roughly 90% of algorithmic control while removing roughly 90% of the infrastructure pain. Research groups at Princeton, Stanford, Berkeley, and Redwood Research were using it before the public launch, and it became the business, with usage-based pricing per million tokens. Note the ordering. The customization infrastructure shipped a full nine months before the company had any model of its own to customize. (Thinking Machines Lab, VentureBeat, Tinker)
Interaction models came second. A May research preview showed AI designed to listen and speak in real time, and even to interrupt, instead of stopping and waiting the way typical chatbots do. This is the interface layer of the thesis: widening the channel between humans and machines rather than deepening machine autonomy. (TechCrunch)
Inkling, released yesterday, completes the argument. It is a Mixture-of-Experts transformer with 975 billion total parameters, 41 billion active, supporting a context window of up to one million tokens, pretrained on 45 trillion tokens of text, images, audio, and video, and released under an Apache 2.0 license, with a preview of a lighter-weight sibling, Inkling-Small, alongside it. The design choices are enterprise choices rather than demo choices. The model is built to give calibrated answers, flagging uncertainty instead of guessing, and it supports controllable thinking effort, letting users trade performance against token spend. On one benchmark the company says it matches the coding performance of Nvidia’s Nemotron 3 Ultra while using a third as many tokens. Critically, it is available for fine-tuning on Tinker from day one. The tool and the model finally meet: a base built to be shaped, and the workshop to shape it in. (Thinking Machines Lab, TechCrunch)
Tooling to customize models. Interfaces to collaborate with them. A model built to be customized. Three products, one sentence.
The Open Questions
The skeptic’s case is not weak, and it deserves to be stated plainly.
First, the economics. An open-weight model earns nothing directly, so the whole business rests on Tinker becoming the toll road, and fine-tuning still requires serious machine-learning talent that most enterprises do not have. The bet only pays if the customization layer gets radically easier. To be fair, chief scientist John Schulman has said that is the plan, describing tools meant to help a wider audience with limited technical knowledge fine-tune models for their own purposes. (TechCrunch)
Second, the accountability question, which may prove the more consequential of the two. Because the company markets Inkling as a starting point for organizations to fine-tune themselves, customers rather than Thinking Machines are responsible for making sure their customizations are safe. The company is unusually candid about this, recommending defense-in-depth, meaning downstream moderation layered around the model rather than reliance on refusals alone, plus human review for outputs used in high-stakes contexts. That is the right disclosure. But a thousand fine-tuned Inklings means a thousand accountability perimeters, distributed exactly as widely as the weights. With the EU AI Act’s next obligations landing August 2, the question of how regulators treat an open base model versus its fine-tuned descendants is about to become operational. (Thinking Machines Lab, TechCrunch)
The Orthogonal Take
Every lab in this industry says it wants AI to benefit everyone and to keep humans in the loop. Thinking Machines is the rare company whose architecture matches the claim: governance built to resist capture, products built for collaboration over autonomy, a model priced at free and built to be given away and reshaped. It survived a co-founder exodus, a stalled mega-round, and an aggressive talent raid without swerving from a thesis it stated on day one. Whether that is conviction or stubbornness will only be distinguishable in hindsight. But in a market where every company is optimizing along the same axis, this is the one that rotated ninety degrees, and after yesterday it finally has the artifact to prove the rotation was real. (TechCrunch, Thinking Machines Lab)