Thinking Machines Lab: Mira Murati, Tinker, Inkling & Funding

A lab built from OpenAI alumni

Thinking Machines Lab was announced in February 2025 by Mira Murati, OpenAI's former chief technology officer, as chief executive. John Schulman, an OpenAI cofounder who had briefly joined Anthropic beforehand, came on as chief scientist, alongside other early researchers and engineers recruited mainly from OpenAI. The company is based in San Francisco. [2]

In July 2025 the company raised a $2 billion seed round at a $12 billion valuation, led by Andreessen Horowitz with participation reported from Nvidia, AMD, Cisco and Jane Street among others. Both figures, a record size for a seed round and a valuation set before any product existed, are the numbers the company's first shipped model would later be measured against. [2][5]

Compute before a product

Thinking Machines Lab's relevance to the field, for its first year and a half, rested more on the scale of resources committed to it than on anything it had released. Its first product, Tinker, is reported to have shipped in October 2025 as an API for fine-tuning open-weight models: a customer submits training jobs and the company runs them on its own infrastructure, later extending support to model families including Kimi, built by Moonshot AI, and Nvidia's Nemotron line. [2]

In March 2026, thirteen months after its founding, NVIDIA committed at least a gigawatt of its unreleased Vera Rubin systems to Thinking Machines Lab, alongside an investment of undisclosed size, under a multiyear partnership. That put compute on the scale of the largest frontier labs behind a company that had, at that point, still not shipped a model of its own. [3]

Atlas interpretation: A record seed round, a gigawatt-scale compute commitment and a founding team drawn from the top of OpenAI's research organization are, on their own, evidence of investor confidence rather than of research output. What makes Thinking Machines Lab worth following is the gap between that scale of backing and what had actually shipped, and whether the eventual releases justified the wait. [3][5]

Two releases that closed the gap

In May 2026, the lab released TML-Interaction-Small as a research preview, a model built around continuous audio, video and text input cut into 200-millisecond segments rather than the usual back-and-forth turn structure, so it can be interrupted and can interrupt back. Thinking Machines described this as a new category of model rather than a feature added to an existing one; that framing is the lab's own, not an independent assessment of how the category will hold up. [4]

Two months later, in July 2026, the lab shipped Inkling, its first open-weights language model, roughly a year after the seed round closed with no product to show for it. Independent analysis at release described Inkling as the strongest open-weights model then published by a US lab, a category Chinese labs including Moonshot AI and DeepSeek had largely held through 2025 and into 2026. Thinking Machines later published a smaller distilled variant as well. [6][7]