Thinking Machines Lab is an American artificial intelligence startup founded in February 2025 by Mira Murati, the former chief technology officer of OpenAI. Based in San Francisco and structured as a public benefit corporation, the company built its team almost entirely out of senior researchers who left OpenAI, Meta AI, Mistral AI, and Anthropic — with a stated goal of making advanced AI models genuinely customizable for outside developers, rather than closed to everyone but the lab that trained them. In July 2026 it released Inkling, its first in-house foundation model, open enough for anyone to download and adapt.
Who is behind it
Murati leads Thinking Machines Lab as CEO. Her co-founders include John Schulman, an OpenAI co-founder who serves as the company’s chief scientist, along with Barrett Zoph, Lilian Weng, Andrew Tulloch, and Luke Metz — all veterans of frontier AI labs. The company raised a $2 billion round in July 2025, led by Andreessen Horowitz, that valued it at $12 billion; backers included Nvidia, AMD, Cisco, Jane Street, and even the government of Albania, Murati’s country of birth, which put in $10 million. In March 2026, Nvidia deepened the relationship with a multi-year deal to supply the lab with a full gigawatt of Vera Rubin computing capacity.
The company’s first product, launched in October 2025, was Tinker — an API that lets outside developers fine-tune open-weight models for their own tasks without having to build or manage the distributed training infrastructure that job normally requires.
What Inkling is
Inkling is the model that gives Tinker something purpose-built to work with. It’s a Mixture-of-Experts transformer with 975 billion total parameters, of which only about 41 billion are active for any given token — each token is routed to 6 of 256 specialized “expert” sub-networks plus two shared ones, which keeps the computing cost of running the model far below what its total size would suggest. It was pretrained on roughly 45 trillion tokens of text, images, audio, and video, reads a context of up to 1 million tokens at once, and can take text, images, and audio as input.
Unlike the flagship models OpenAI, Anthropic, and Google keep behind an API, Thinking Machines Lab released Inkling’s full weights under an Apache 2.0 license, so anyone can download, modify, and even commercialize it. The company also previewed a smaller, cheaper companion model, Inkling-Small. The pitch isn’t that Inkling tops every leaderboard — Thinking Machines Lab has said explicitly that it optimized for a strong, adaptable base for fine-tuning rather than for raw benchmark scores. Developers can get the weights directly from Thinking Machines Lab’s official Hugging Face page, which also lists deployment options through several hosting providers.
Why it matters
Inkling puts Thinking Machines Lab in an unusual position: a well-funded lab stacked with former OpenAI leadership choosing to give away its core model rather than sell access to it. That decision puts it in direct competition with the wave of strong open-weight models coming out of Chinese labs — DeepSeek, Alibaba’s Qwen, and Zhipu AI’s GLM among them — which have been winning developers over on price and openness even as most US labs kept their best models closed. It also signals what Murati has described as the company’s real bet: that the next phase of AI competition is less about who has the single smartest model, and more about who lets businesses adapt a model to their own data and workflows.
In the news
Thinking Machines Lab’s release made today’s briefs: Thinking Machines Lab Releases Inkling, an Open-Weight AI Model.
FAQ
Is Thinking Machines Lab connected to OpenAI? Not officially. It’s an independent company, but nearly its entire founding team, starting with CEO Mira Murati, previously worked at OpenAI.
Is Inkling free to use? The weights are free to download and modify under an Apache 2.0 license; running the model at scale still requires substantial computing hardware, or a paid hosting provider.
What’s the difference between Inkling and Tinker? Inkling is the AI model itself. Tinker is the separate platform Thinking Machines Lab built for fine-tuning open-weight models, including Inkling.
Sources: Thinking Machines Lab’s Inkling announcement and model card; Wikipedia.