Why do Chinese labs lead US labs in releasing capable open-weights models?

Anjney Midha · Apr 14, 2026

Chinese labs catch up by doing adversarial distillation of Western models at scale, releasing the results as open models to harvest feedback, and will stop open sourcing once they reach the frontier

Open source functions as a bootstrapping mechanism: distill, release, get feedback, iterate; once good enough for domestic needs there is no reason to keep releasing

38:14 20VC: Anj Midha on Investing $300M into Anthropic | The Early Days of Anthropic & How 21 of 22 VCs Turned it Down | The Four Bottlenecks to Compute | What the China Has Smashed and Why We Should Be Worried

Clay Bavor · Jul 4, 2026 · hedged

Chinese labs have closed much of the gap partly because they are willing to do scaled distillation of frontier models rather than build frontier models themselves

If you can't build frontier models yourself, the next best approach is to distill them and offer them up

Scope: stated as 'probably' a partial driver of the difference

0:00 20VC: Open Models vs Frontier Models: Who Actually Wins? | The $100,000 Token Budget Every Engineer Will Need | Why Forward-Deployed Engineers Are the Future of Enterprise AI with Clay Bavor, Co-Founder of Sierra

Clay Bavor · Jul 4, 2026

US labs and hyperscalers rationally won't release open weights models of similar capability to their frontier models because it would cannibalize and price-pressure their own frontier business, and this is the main driver of the US–China open model gap

If you build frontier models you would not compete with yourself; if you can't build frontier models, distilling and releasing them is your next best option

17:35 Us labs withhold to avoid cannibalizing frontier revenue

Anastasios Angelopoulos · Aug 3, 2026

Kimi's result was a genuinely big moment because it falsifies the persistent US narrative that Chinese labs only keep up by distilling American models.

Kimi beat all American models including Fable on some subset of tasks; distillation may still be a sub-step, but the labs must be doing something beyond distillation to exceed American performance.

6:23 Chinese labs exceed american models so more than distillation

Anastasios Angelopoulos · Aug 3, 2026

US open source has lagged because nobody had figured out a business model for it, and there are now two viable ones: revenue share with inference providers, and using the open model as lead generation for forward-deployed engineering work.

Rev share lets you capture part of the compute revenue; the lead-gen route feeds into AI modernization work — retooling every business in the world around AI — which will be one of the biggest markets of the next decade.

11:01 Us lag is a missing business model not missing capability