SemiAnalysis counts 857 models from nine leading Chinese AI developers: only 9 shipped with a published safety evaluation, and 94.9% have no safety disclosure at all

Why it matters

【Key points】

SemiAnalysis published "Beijing Will Not Pace the Frontier" on Oct 8, built on a census of 857 model releases from nine Chinese developers (ByteDance, Alibaba, Tencent, Baidu, DeepSeek, Moonshot, Zhipu, MiniMax, StepFun), 2021 to Sep 15, 2026.

Only 31 (3.6%) ever came with a developer-published safety result for that model, and just 9 (1.1%) had it at launch; 16 were documented later (median lag 42 days). 813 (94.9%) have no safety disclosure. Startups disclosed for 6.3% of releases versus 2% for the four hyperscalers. Among 2026 flagships only Zhipu's GLM-5.3 shipped with an evaluation note.

The piece argues China regulates content and applications strictly but has no compute- or capability-triggered frontier duties, looser than the EU and California's SB 53.

【Verifiable】

Article (partly paywalled): https://newsletter.semianalysis.com/p/beijing-will-not-pace-the-frontier

SemiAnalysis post on X: https://x.com/SemiAnalysis_/status/2108344060658221528

Caveat: "not found" does not mean "not tested", and per-company rates are indicative. About 115 posts on X in 24 hours in English and Chinese, steady through the day.

【Our read】

1. This measures disclosure, not safety, but for open weights disclosure matters more: once weights ship, downstream users only have the launch documents. More than half the releases are open-weight, and under 1% had safety results at launch.

2. The gap has already traveled abroad. Microsoft-Decision-1 and Cloudflare's Clef models are post-trained on Qwen weights and sold to US enterprises, so due diligence on Chinese model safety now falls on Western clouds and buyers.

3. Watch the next flagship launches, not statements: whether the next GLM, DeepSeek, Qwen and Kimi model cards ship with concrete cyber, bio and jailbreak numbers on day one. Zhipu has started; whether others follow will say more than any framework text.