Thinking Machines
Why it matters
Deep Dive
Thinking Machines Lab was founded in early 2025 by Mira Murati, who spent six and a half years at OpenAI and served as its chief technology officer during the launches of ChatGPT and GPT-4. The founding team reads like a roll call of modern LLM development: Barrett Zoph, a research scientist behind much of ChatGPT's post-training, became CTO, and John Schulman, an OpenAI co-founder who helped pioneer RLHF, joined as chief scientist. The lab's stated goal is to close the gap between fast-advancing AI capabilities and the broader community's ability to understand, customize, and actually use those systems, with an emphasis on human-AI collaboration over full automation. In 2025 it shipped its first product, Tinker, a managed service for fine-tuning open-weight models, and it has committed to publishing research as it goes.
The Founding Team
Talent is the whole story here. Murati left OpenAI in late 2024 after briefly serving as interim CEO during the November 2023 leadership crisis; within months she had assembled a founding team drawn heavily from OpenAI, with additional researchers from Anthropic, Google DeepMind, and Meta. Zoph co-led the post-training work that turned raw GPT-4 into ChatGPT, and Schulman was a lead author of both the PPO algorithm and the RLHF pipeline that became the standard way to align chat models. The pattern is familiar — Anthropic itself was founded by OpenAI alumni in 2021, and Safe Superintelligence was started by another OpenAI co-founder — but Thinking Machines assembled arguably the densest concentration of post-training expertise outside a frontier lab.
Tinker: Fine-Tuning as an API
Tinker, the lab's first product, is a fine-tuning service exposed as an API, but with an unusual level of control. Most fine-tuning offerings are black boxes: you upload a dataset and get back a model. Tinker instead exposes low-level training primitives — forward and backward passes, optimizer steps, sampling from the current weights — so the user writes the training loop in ordinary Python while the lab handles scheduling the work across its GPU clusters. The service fine-tunes open-weight models, including the Llama and Qwen families, and it uses LoRA so that many customers can share the same base weights while each keeps a small, cheap adapter. That design keeps both supervised fine-tuning and reinforcement learning workflows practical without the user renting or configuring a single server.
Not a Frontier Model Lab (Yet)
Given the valuation, the natural assumption is that Thinking Machines is racing OpenAI and Google to train a bigger foundation model. So far it is not: the company has released no pre-trained model of its own, and its first product builds on other labs' open weights rather than competing with them. The implicit bet is that the industry has over-invested in pre-training scale and under-invested in post-training — the reinforcement learning and fine-tuning stage where models actually become useful for a specific task — and that giving outside researchers and companies real control over that stage is a business in itself. Whether that bet holds is an open question, but it is a deliberate positioning choice, not a failure to ship.
The $12 Billion Bet
In mid-2025 the company raised roughly $2 billion in a seed round that valued it at about $12 billion — one of the largest seed financings in tech history, closed before the company had announced any product. The number reflects a specific investor thesis from the post-2023 talent wars: that a small group of researchers with frontier-lab experience is worth more than almost any other asset in AI. In practice the capital buys two things the lab cannot get otherwise — large reservations of GPU compute and the compensation packages needed to hold a team that every rival is trying to poach. It also sets an unusually high bar: at that valuation, Tinker needs to grow into a real platform business, and the research agenda needs to justify the lab standing alongside the companies its founders left.