Thinking Machines Lab

Mira Murati left OpenAI in 2024 after a stint as chief technology officer, and by February 2025 she had her own shop: Thinking Machines Lab, a San Francisco public benefit corporation built around the idea that AI research should not be locked behind a handful of corporate walls. The company's stated mission is "a future where everyone has access to the knowledge and tools" to work with AI on their own terms. It is deliberately small, listed on LinkedIn at 11 to 50 employees, and headquartered at 2300 Harrison St in San Francisco.

The first public product was Tinker, an API for fine-tuning open-weight models that launched in October 2025. Tinker lets developers adapt families like Inkling, Qwen, and Nemotron, plus models such as DeepSeek-V3.1, GLM-5.3, Kimi K2.6, and gpt-oss, without building the training stack from scratch. The flagship in-house model, Inkling, arrived on July 15, 2026, under the Apache License. At 975 billion parameters it was the largest and most capable open-weights model released outside China at the time, edging past Nvidia's Nemotron 3 Ultra. A distilled 276-billion-parameter Inkling Small shipped the same month with performance close to the full model. Thinking Machines disclosed that Inkling was built in part on Chinese open-weights foundations, using DeepSeek-V3 as the base and Moonshot AI's Kimi K2.5 for post-training synthetic data, a detail that drew attention from policy watchers in Washington.

The money arrived fast. By July 2025, Thinking Machines had closed a $2 billion early-stage round led by Andreessen Horowitz at a $12 billion valuation, with Nvidia, AMD, Cisco, and Jane Street participating. That kind of check for a company with no shipped model raised eyebrows, but it bought time and compute. By September 2026 the company was seeking roughly $1 billion at a $40 billion valuation, with Accel reportedly in talks to lead the round, while Nvidia was said to be negotiating an investment of around $2.5 billion at that same mark. On the revenue side, the company reached a $100 million annualized run rate on Tinker by October 2026, a meaningful inflection for a firm barely 20 months old. Venture capitalists have nonetheless begun publicly questioning "neolab math," i.e., whether these valuations are justified by underlying economics.

Infrastructure deals suggest the company is buying compute and inference capacity rather than building its own data centers. In September 2026 it struck a $65 million inference partnership with Crusoe, and in October 2026 Australian data center firm Firmus signed Thinking Machines as its first major contract following its failed IPO. The brand leans on Murati's personal credibility and an anti-closed, pro-accessibility message. Its visual identity is minimal and typographic, and its marketing is unusually research-forward: model cards, license disclosures, grant programs, and technical transparency rather than consumer advertising. The company communicates primarily through its website, its X account at @thinkymachines, and technical releases. [edit]