Thinking Machines Lab

What it is

Thinking Machines Lab is an artificial intelligence company based in San Francisco, founded in February 2025 and organised as a public benefit corporation — a US corporate form in which a company commits to a stated public benefit alongside its commercial goals.^[1][4]

Origins and founding team

The company was founded by Mira Murati, who had been chief technology officer at OpenAI.^[1][4] Its founding team included Barret Zoph, Lilian Weng and Luke Metz.^[1] John Schulman became chief scientist after a brief stint at Anthropic, and the company's advisers include Bob McGrew and Alec Radford.^[1]

In July 2025 Thinking Machines Lab raised about $2 billion at a $12 billion valuation in an early-stage round led by Andreessen Horowitz; Nvidia, AMD, Cisco and Jane Street were reported among the investors.^[1] Albania's government also made a $10 million investment, which required an amendment to the country's 2025 budget.^[1]

Tinker and fine-tuning

On 1 October 2025 the company launched Tinker, an API for fine-tuning open-weight models.^[2] Fine-tuning means taking a model that has already been trained and continuing to train it on additional data so that it behaves differently on some task; "open-weight" means the trained parameters are published, so the work can be done on the model itself. Tinker exposes low-level primitives such as forward_backward and sample, from which most common post-training methods can be implemented.^[2] It supports open-weight models of several sizes, including large mixture-of-experts models such as Qwen-235B-A22B.^[2] Tinker ran in private beta for researchers and developers and was initially free of charge.^[2]

Research and the first model

The company publishes research under the Connectionism blog, including posts on LoRA, on-policy distillation and defeating nondeterminism in LLM inference.^[1] On 15 July 2026 it released its first model, the open-weight Inkling, with 975 billion parameters, followed by Inkling-Small.^[3] Inkling's full weights, including an NVFP4 version, were published on Hugging Face.^[3]

People and partnerships

Andrew Tulloch left for Meta in October 2025, and Barret Zoph and Luke Metz returned to OpenAI in January 2026.^[1] In March 2026 Nvidia announced a strategic partnership with the company, with an undisclosed investment and a multi-year deal to deploy one gigawatt of Vera Rubin compute.^[1] In August 2026 the company moved to 2300 Harrison Street in San Francisco's Mission District.^[1]

Where it sits

Thinking Machines Lab entered the field with a founding group drawn largely from OpenAI, and with unusually large early funding for a new company: a $2 billion round at a $12 billion valuation within months of being founded.^[1] Its public output has centred on open-weight models and on the tooling for adapting them: Tinker gives researchers and developers direct access to training primitives for fine-tuning published models, and Inkling was released with its full weights.^[2][3]

The company has also kept a public research presence through its Connectionism blog, writing about adaptation methods such as LoRA and on-policy distillation and about practical problems such as nondeterminism in inference.^[1] Its compute agreement with Nvidia — a multi-year deal for one gigawatt of Vera Rubin compute — points to the scale of infrastructure involved in training and serving models of Inkling's size.^[1][3] Staff have moved in both directions between the company and other AI labs, and the company relocated within San Francisco in 2026.^[1]