Thinking Machines Lab has unveiled its first general‑purpose AI model, the Inkling foundation model, built in San Francisco. The lab says its architecture draws on China’s Deep Seek‑V3 and was refined post‑training with data from Beijing‑based Moonshot AI’s Kimi K2.5.I. The release comes as Chinese models have overtaken U.S. counterparts in overall usage this year, according to Open Router data.

Unlike the closed‑weight GPT or Claude models, Inkling will be open‑weight, allowing users to host and fine‑tune the full model on their own servers. The company claims it trained the model from scratch, making it fully available for customization. Inkling ranks below leading U.S. and Chinese models in benchmarks, but the lab argues it offers greater flexibility and user agency.

The start‑up raised a $2bn seed round at a $12bn post‑money valuation, backed by Andreessen Horowitz, Nvidia, AMD and Jane Street. Mira Murati, former Open AI CTO who helped build Chat GPT and DALL‑E, left Open AI in September 2024 to found the lab in February 2024. The team also launched the Tinker platform, enabling enterprises to fine‑tune large‑language models for specific applications.