Mira Murati’s startup released a 975 billion-parameter multimodal model under an Apache 2.0 license, adding Tinker tooling for customization after a record $2 billion seed round.
Thinking Machines Lab has released Inkling, the first public AI model from the startup founded by former OpenAI executive Mira Murati, expanding the launch with technical details that position it as one of the largest open-weight systems to date. The model has 975 billion parameters, uses a Mixture-of-Experts architecture with 41 billion parameters active at a time, supports text, images, audio and video, and comes with a 1 million-token context window. Its weights are available on Hugging Face under the Apache 2.0 license, and the company is also launching a customization platform called Tinker. The release builds on Thinking Machines’ rapid rise since its February 2025 founding. The company raised $2 billion at a $12 billion valuation in what research suggests was the largest seed round in venture capital history, with backers including Andreessen Horowitz, Nvidia, AMD, Cisco and Jane Street. Reports suggest ongoing discussions could push its valuation toward $50 billion. Nvidia also signed a multiyear chip supply agreement in March 2026 to provide Vera Rubin accelerators. Inkling was trained from scratch on NVIDIA GB300 NVL72 hardware using 45 trillion tokens. Benchmark results including SWE-bench Verified indicate it is competitive, though not the clear leader in every category. The launch also follows key staff departures in January 2026 that had raised questions about stability, with the release signaling the company can still ship a model of significant scale on an aggressive roughly 18-month timetable. The broader significance lies in how open-weight deployment can let developers fine-tune and run advanced models on their own infrastructure, while potentially creating downstream demand for decentralized GPU compute networks and derivative models tailored to areas such as trading, DeFi risk analysis and smart contract auditing.