SemiAnalysis: 3.6% of 857 Chinese AI Releases Had Safety Evals

Of 857 model releases from nine leading Chinese AI developers between 2021 and September 2026, 31 came with a published safety evaluation. Nine had results available at launch. The rest, 813 releases or 94.9%, shipped with no safety disclosure at all. That is the headline number from a new SemiAnalysis report covering the hyperscalers ByteDance, Alibaba, Tencent and Baidu alongside the startups DeepSeek, MoonShot, Zhipu Z.ai, MiniMax and StepFun.

The authors are blunt about what the data shows: "China's real approach to AI safety is speed-based, not safety-based. Though Beijing recognizes frontier AI risks, it has been prioritizing rapid development." They point at the AI Safety Governance Framework 3.0, published September 14, 2026, which warns of a "self-accelerating trend of autonomous learning" and then opens its own text with "promoting AI innovation and development as the first priority."

The disclosure that does happen tends to arrive late. The median lag between a model release and a published safety evaluation is 42 days; the maximum in the sample is 349. For reasoning models specifically, the report finds 93% have no published safety results at all.

Individual leaders show up differently in the record. Zhipu CEO Zhang Peng is singled out as the only one who has consistently prioritized safety across international commitments. DeepSeek founder Liang Wenfeng, by SemiAnalysis's reading, has made no public safety statements in four years. Alibaba's Eddie Wu delivered what the report describes as a "6,000-word keynote on the road to superintelligence in September 2025 that does not contain the word safety or risk once and closed with 'we are full of optimism.'" Alibaba's own security vice-president Xue Hui is quoted conceding that AI safety capability "currently lags the development of model capability."

The regulatory architecture reinforces the pattern. SemiAnalysis lists mandatory content labeling (September 2025), Cybersecurity Law Article 20 (January 2026), minors' information classification rules (March 2026), MIIT's AI ethics-review regime (March 2026), policy-level agent rules (May 2026), AI-companion rules (July 2026) and network-data risk assessment rules (August 2026), and then notes: "There is no set of frontier-risk duties triggered by training compute or model capability." Content and applications are regulated; the models themselves, before deployment, are not.

The analysis arrives into a dense beat: this is one of 333 China-AI stories we have tracked in the last 90 days, and sits next to a parallel run of 325 safety stories over the same window. Xi Jinping has publicly said "the faster AI advances, the more promptly the measures to prevent loss of control must be improved." The 3.6% number is what that sentence looks like once you count the releases.