Teams that want to fine-tune and own a model, rather than rent one, are the natural buyers of Thinking Machines Lab. Wikipedia counts about 100 staff and NVIDIA has committed at least one gigawatt of compute, yet the company itself concedes Inkling is not the top model on the market.
Thinking Machines Lab is a San Francisco AI company led by former OpenAI chief technology officer Mira Murati. It sells the Tinker fine-tuning API and trains the Inkling model family, whose newest member, Inkling 2 Small, has 594 billion total parameters.
Founded: 2025 · HQ: San Francisco, CA, USA · Team: 50-200 · CEO: Mira Murati · Funding: $2B seed round in July 2025 (Andreessen Horowitz lead); a further NVIDIA investment in March 2026, amount undisclosed; a $5B round led by Accel is reported but unconfirmed · Valuation: $12B at the seed round; a $40B pre-money valuation is reported but unconfirmed
About Thinking Machines Lab
Thinking Machines Lab was founded in February 2025 in San Francisco by Mira Murati, who served as chief technology officer of OpenAI until September 2024. Her co-founders included Barret Zoph, formerly OpenAI's vice president of research, and Luke Metz. John Schulman, an OpenAI co-founder who spent a short spell at Anthropic in 2024, joined as chief scientist at launch. The early team was recruited from OpenAI, Meta AI and Mistral AI, and the company's trust center describes it as an artificial intelligence research and product company.
The first product was Tinker, a fine-tuning API announced in October 2025. Customers write their own training loop and loss, while Thinking Machines runs the GPUs for sampling and training and stores checkpoints. The company says users have trained custom models for financial judgment, hardware design, search agents and specialized coding, and that customer content on Tinker belongs to the customer and is never used for training. The in-house models followed. Inkling, a mixture-of-experts model with 975 billion total and 41 billion active parameters and a 1M-token context window, shipped as an open-weights release on 15 July 2026, and the smaller Inkling-Small had its weights out by 31 July.
The second generation arrived on 5 November 2026 with Inkling 2 Small, a 594 billion parameter model with 17 billion active that reads text and images. Artificial Analysis scores it 31 on its Intelligence Index. Unlike the first generation, it launched on Tinker with closed weights. The company says it intends to publish them under Apache 2.0 after external safety testing, and its Hugging Face repository was still not public on 9 November 2026.
In March 2026 the company announced a multi-year partnership with NVIDIA to deploy at least one gigawatt of Vera Rubin systems, with deployment targeted for early 2027, and NVIDIA made an undisclosed investment. Leadership changed in January 2026, when Murati announced that CTO Barret Zoph had left. Soumith Chintala was named CTO, and Zoph, Metz and Sam Schoenholz returned to OpenAI. Co-founder Andrew Tulloch had already left for Meta in October 2025. Wikipedia puts the headcount near 100 in 2026.
The stated mission is to build AI that extends human will and judgment, and the company publishes research on its Connectionism blog. A July 2026 post, A Safe Path to Open Weights, describes a staged approach to releases. Before publishing Inkling it ran internal evaluations, a fine-tuning study to elicit worst-case capabilities, and external red-teaming with Scale AI, Handshake AI, FAR.AI and Apollo Research. It held back the Inkling 2 Small weights pending further safety evaluations.
The company wins where customers want to own and adapt a model. Tinker also hosts third-party open-weight models such as Qwen, Nemotron, GLM-5.3, Kimi K2.6 and gpt-oss, and the company reports the best open-weight finance results on the benchmarks it published for Inkling 2 Small. It loses on peak capability: its own July 2026 post says Inkling is not the strongest model available, and its model card shows Claude Haiku 5.5 ahead on the Artificial Analysis index. Buyers weighing open-weight rivals can compare Kimi K3 from Moonshot AI or GLM-5.3 Flash from Z.ai.
Mission
Build AI that extends human will and judgment.
Products
- Tinker (Fine-tuning and inference API): https://thinkingmachines.ai/tinker/
- Inkling 2 Small (Multimodal MoE model): https://thinkingmachines.ai/news/inkling-2-small/
- Inkling (Open-weights flagship model): https://thinkingmachines.ai/news/introducing-inkling/
- Inkling-Small (Open-weights MoE model): https://huggingface.co/thinkingmachines/Inkling-Small
- Tinker Playground (Browser chat and sampling console): https://tinker.thinkingmachines.ai/playground
Thinking Machines Lab Models on Hokai
Compliance
SOC 2 Type I
Links
Frequently Asked Questions
How much has Thinking Machines Lab raised, and at what valuation?
Its $2 billion seed round closed in July 2025 at a $12 billion valuation, led by Andreessen Horowitz with Accel, NVIDIA, AMD and Jane Street, according to TechCrunch. NVIDIA added an undisclosed investment alongside its March 2026 compute partnership. Dataconomy reported on 6 November 2026 that a $5 billion round led by Accel at a $40 billion pre-money valuation could close within the week; the company declined to comment, so treat that figure as unconfirmed.
What does Thinking Machines Lab sell, and what does it cost?
Tinker is the commercial product: an API for fine-tuning and sampling Inkling and third-party open models, billed per million tokens. Inkling 2 Small lists at $0.30 prefill, $0.03 cached prefill and $1.00 sampling on Tinker, with training at $1.50 for the standard context. The earlier Inkling and Inkling-Small models are also on Tinker, and their weights are on Hugging Face.
Does Thinking Machines Lab train on customer data, and is it SOC 2 certified?
The Tinker terms say customer content, including datasets, prompts and outputs, is not used to train or fine-tune models or shared with third parties for that purpose, and it is retained only as needed to run the service. The trust center lists a SOC 2 Type I report available on request plus a Tinker penetration test; we found no SOC 2 Type II, ISO 27001 or HIPAA claim. Named subprocessors include Google Cloud Platform in the United States, Cloudflare, WorkOS and OpenAI's Moderation API.
Who does Thinking Machines Lab compete with?
OpenAI and Anthropic sell closed models with higher peak capability, and Thinking Machines positions Tinker's customer-owned fine-tuning against them. Among open-weight labs, Moonshot AI, Z.ai and Mistral AI make models that its own comparison table sets against Inkling 2 Small. NVIDIA is both a rival through its Nemotron models and the supplier of its compute.
How do you start using Thinking Machines Lab's models?
Sign up for Tinker, then call Inkling 2 Small from the Tinker Playground or the SDK using the model ID thinkingmachines/Inkling-2-Small. The Tinker Cookbook on GitHub holds fine-tuning recipes, and the tml-renderers package on PyPI handles tool calls and image inputs. Serverless inference is in beta behind a waitlist, so production traffic needs an email to the Tinker team first.