Sequoia: Qwen Went From 1% to 69% of All Open-Model Fine-Tunes in Two Years — American Startups Are Building on Chinese Weights

Sequoia Capital·high signal

Sequoia published 'America's Open-Model Paradox' on July 24, arguing Western AI companies have no legal path to distill their own frontier models, so they post-train on Chinese open weights instead. Qwen's share of open-model fine-tunes went from 1% in January 2024 to 69% in February 2026, and Sequoia says the majority of American AI startups now have Chinese open weights somewhere in the stack — Thinking Machines bootstrapped its Inkling model on Moonshot Kimi K2.5 synthetic data. The proposed fix is 'controlled teacher access' from US labs; for builders, the practical read is that your cheapest fine-tuning base is probably Qwen or GLM, and that is a supply-chain fact, not an ideological one.