Chinese AI Labs Publish Safety Results for Only 3.6% of 857 Model Versions Since 2021

The SemiAnalysis investigation, entitled “Beijing Will Not Pace the Frontier,” assembled a thorough catalogue of artificial‑intelligence model releases originating from nine prominent Chinese developers over the period spanning 2021 up to 15 September 2026. In total, the researchers identified 857 distinct model versions, thereby establishing a robust baseline against which safety‑related disclosures could be measured.

Out of those 857 versions, merely 31 – exactly 3.6 % – featured a safety result that the respective developer elected to make public. This proportion is derived directly from the SemiAnalysis calculations and is presented as an established fact without alteration.

Even more striking is the timing dimension: only nine of the 31 disclosed results, representing 1.1 % of the total releases, were made available concurrently with the model’s launch. The remaining 22 safety results were published only after the model had already become accessible to the public.

Distribution of disclosures across company types

When the dataset is segmented by organisational category, a clear divergence emerges. Start‑up firms such as DeepSeek, Moonshot, Zhipu, MiniMax and StepFun display a higher propensity to share safety outcomes – 6.3 % of their releases include a published result – whereas the four hyperscale giants (Alibaba, ByteDance, Tencent and Baidu) together disclose merely 2 % of theirs.

Among the nine entities, Zhipu – operating under the Z.ai brand – stands out as the sole organization that has consistently published a safety result every year since 2022, according to statements released by the company itself.

Absolute numbers for the biggest players

Alibaba, which holds the largest catalogue of releases with 238 distinct versions, contributed only seven safety results to the public record. Tencent, with 133 versions, released a single safety assessment, while ByteDance disclosed two safety results out of its 120 releases. These figures are reported by the companies themselves and reflect the limited transparency observed.

The remaining 813 versions – accounting for 94.9 % of the total model count – have no public safety disclosure whatsoever. This proportion is presented as an established fact in the SemiAnalysis report and underscores the scale of the transparency gap.

Timing of post‑launch disclosures

Among the 31 disclosed safety results, sixteen were published after the corresponding model had already been launched. The median delay between launch and disclosure was 42 days, while the longest lag stretched to 349 days. These timing statistics are taken directly from the study’s analysis.

• Only 3.6 % of Chinese AI model releases include any public safety result.
• Startups disclose safety results at a rate of 6.3 %, hyperscalers at 2 %.
• Zhipu has published a safety result every year since 2022.
• Median post‑launch disclosure delay is 42 days.

The “China AI Safety Governance Framework 3.0”, which was made public on 14 September 2026, recommends regular testing and periodic public disclosure of safety evaluations. However, the framework does not impose a legal requirement to comply, meaning that adherence remains entirely voluntary for developers.

The authors of the SemiAnalysis study caution that the absence of public disclosures does not necessarily imply that internal safety testing is missing. Their methodology captures only information that is publicly accessible, leaving a potential blind spot for evaluations that remain private.

For English‑speaking organisations that are evaluating Chinese AI providers, the practical implication is clear: without mandatory disclosure standards, companies must assume that the majority of model releases have undergone limited or undocumented safety scrutiny. Procurement teams should therefore request internal safety audit reports, negotiate contractual clauses that require timely disclosure of safety results, and consider third‑party verification to mitigate the risk of undisclosed vulnerabilities.

In light of these findings, industry analysts suggest that investors and regulators monitor the evolving transparency practices of Chinese AI labs closely, as any shift toward more systematic disclosure could materially affect risk assessments and market confidence.

The SemiAnalysis team also notes that the modest increase in disclosures from startups may reflect competitive pressures to demonstrate responsible AI practices, whereas hyperscalers appear to prioritize rapid rollout over public safety reporting.

Despite the modest progress observed, the overall landscape remains one of limited openness, with nearly nine out of ten model releases lacking any publicly available safety assessment, a statistic that underscores the need for stronger governance mechanisms.

Stakeholders are encouraged to engage with Chinese developers directly to obtain detailed safety documentation, especially for high‑impact models that could pose significant societal risks if inadequately vetted.

Finally, the SemiAnalysis report recommends that future research incorporate confidential surveys of developers to capture non‑public safety testing activities, thereby providing a more comprehensive picture of the true state of AI safety in China.