SemiAnalysis reports Chinese AI labs rarely publish safety results

The core of the report is a systematic dataset of 857 model releases from nine major Chinese developers (including ByteDance, Alibaba, DeepSeek, and Zhipu) between 2021 and September 2026. The analysis found that only 3.6% of these releases were accompanied by any published safety evaluation from the developer. Just 1.1% had results available at or before the model’s launch. The remaining 94.9% shipped with no safety disclosure at all. This trend has accelerated as release cadence has increased thirty-fold since 2023, with open-weight models making up more than half of recent releases.

The report also reviewed 65 public statements from lab leaders, finding that safety engagement is sparse and concentrated primarily in Zhipu (Z.ai). Other major players, including hyperscalers like Alibaba and Tencent, have rarely published quantitative safety findings for their frontier models. SemiAnalysis concludes that Chinese labs are in a release and price war, and that current compliance mechanisms do not ensure frontier safety, as a lab can meet all regulatory requirements without ever running dangerous-capability evaluations.