Thinking Machines Lab

Thinking Machines Lab is an American artificial intelligence company founded in 2025 by Mira Murati, who served as chief technology officer of OpenAI from 2022 until her departure in September 2024. The laboratory builds general-purpose AI systems and developer tooling, and became known for shipping both a managed fine-tuning service and an open-weight mixture-of-experts model within its first two years.^[1]

History

Murati announced Thinking Machines Lab in mid-2025, recruiting researchers and engineers from OpenAI, Google DeepMind and other frontier laboratories. In July 2025 the company closed a $2 billion seed round led by Andreessen Horowitz at a $12 billion valuation — among the largest seed financings in Silicon Valley history — before it had shipped a public product.^[2]

The first product, Tinker, launched in October 2025 as an API-based platform for customising and fine-tuning large language models without requiring teams to manage training infrastructure.^[3] In July 2026 the company released Inkling, its first general-purpose model, as open weights. Later in 2026, reports placed the company in discussions to raise additional funding at a valuation of at least $40 billion.^[1]

Technology

Inkling uses a mixture-of-experts (MoE) architecture with 975 billion total parameters of which roughly 41 billion are active per token, a context window of up to 1 million tokens, and pretraining across 45 trillion tokens of text, images, audio and video. The company reports training it specifically for coding, tool use and multimodal tasks; it scored 77.6% on SWE-Bench Verified in the lab's own evaluation and exposes a tunable reasoning-effort setting (0.2–0.99) that trades output quality against token count.^[3]

Because the weights are open, Inkling can be self-hosted with common inference stacks such as vLLM, SGLang and llama.cpp, and is also served through API platforms including Together AI, Fireworks, Modal, Databricks and Baseten. Developers can fine-tune it through Tinker at context lengths of 64,000 or 256,000 tokens.^[3]

Why It Matters for AI

Thinking Machines Lab is the highest-profile example of the frontier-lab spin-out pattern — a senior team leaving an established laboratory to raise consumer-scale capital independently. Its open-weight release strategy also positions it as a United States answer to open-weight models from Chinese developers such as DeepSeek and Zhipu, which have led parts of the open model leaderboard.^[1] The Tinker platform reflects a broader shift of fine-tuning from a research exercise into a managed commercial service.

>See Also

• OpenAI
• Open-Weight Models
• Fine-Tuning
• Mixture of Experts
• Anthropic