# China's Nine Leading AI Developers Published Model-Specific Safety Results for 3.6% of 857 Releases, SemiAnalysis Finds

Times of AI Desk · 2026-10-09 · Research

[https://timesof.ai/2026/10/semianalysis-china-ai-safety-disclosure-3-6pct](https://timesof.ai/2026/10/semianalysis-china-ai-safety-disclosure-3-6pct)

> SemiAnalysis reviewed 857 model releases from 2021 to Sept. 15, 2026 by nine leading Chinese developers and found 31 (3.6%) with a published safety-evaluation result tied to that specific model, and nine (1.1%) with results available at or before launch. It found no safety disclosure for 813 releases, which does not mean they were untested. The report gives no comparable figure for U.S. developers, so it measures disclosure in China, not a gap with the U.S.

China's leading AI developers ship models far faster than they publish safety results for them, on SemiAnalysis's count. The research firm reviewed **857** releases by nine Chinese developers from 2021 to **September 15, 2026** and found that only **31**, or **3.6%**, have ever been accompanied by a published safety-evaluation result from the developer that can be matched to that model. Just **nine** (**1.1%**) had the result available at or before release. The report, by Mark Chen, Doug and Dylan Patel, is dated October 8; Reuters reported it on October 9.

What it measures is disclosure. SemiAnalysis says "not found" is bounded to the materials it checked "and does not mean 'not tested.'"

## How SemiAnalysis counted

- **Who:** four large tech groups (**ByteDance**, **Alibaba**, **Tencent**, **Baidu**) and five startups (**DeepSeek**, **Moonshot**, **Zhipu (Z.ai)**, **MiniMax**, **StepFun**).
- **What:** 857 counted releases, 741 product models and 116 research models, each checked against the developer's model cards, release notes and technical reports.
- **The bar:** a quantitative or substantive finding on harmful output, jailbreaks, toxicity, privacy, refusal or dangerous capability, tied to the named model. Statements that a model was "safety-trained" or "evaluated" do not count, and a flagship's result is not extended to other sizes or snapshots.

Because companies name variants at different granularity (Alibaba's 238 releases count every Qwen size and snapshot), SemiAnalysis calls its per-company rates "indicative rather than a ranking."

## What it found

| Measure | Releases | Share |
| --- | --- | --- |
| Any published model-specific safety result | 31 | 3.6% |
| Result available at or before launch | 9 | 1.1% |
| No safety disclosure | 813 | 94.9% |
| Startups with any result | 20 of 317 | 6.3% |
| Large tech groups with any result | 11 of 540 | 2% |

Of the remaining releases with results, 16 were documented only after launch, with a median lag of 42 days, and for six the timing or model match could not be established. SemiAnalysis says Zhipu is the only developer with a result every year since 2022, and that every 2026 frontier release it lists was undocumented at launch except GLM-5.3, which came with a capability evaluation note. It says no Chinese frontier text model has shipped with a dangerous-capability evaluation across the domains named in the IDAIS statements; Reuters renders this as no major Chinese developer having released a frontier text model with public dangerous-capability tests spanning cyber, biological and loss-of-control risks.

SemiAnalysis also argues that Beijing's binding rules govern applications and their effects on users rather than requiring frontier developers to run or publish capability-based risk assessments. That is the firm's reading of the rules, not a finding of the count.

## What the report does not show

Reuters notes that the report did not provide comparable figures for U.S. developers, adding that leading U.S. companies including OpenAI, Anthropic and Google DeepMind have published safety reports, system cards or model cards for some major frontier launches. Without a matched U.S. count under the same rules, the 3.6% cannot be read as a measured gap between the two countries.

## Limits

- SemiAnalysis's report is partly paywalled; the free portion, which carries the dataset results above, was read in full and the paid section was not.
- Reuters (read via CNBC-TV18) reports the same study, so it is not independent verification of the numbers.
- None of the nine companies' responses to the report has been published that we found.

## Sources

- [SemiAnalysis: Beijing Will Not Pace the Frontier: China's Speed-First AI Safety Regime (October 8, 2026; free portion only)](https://newsletter.semianalysis.com/p/beijing-will-not-pace-the-frontier)
- [CNBC-TV18 (Reuters): China AI developers publish safety tests for just 3.6% of model releases, report finds (October 9, 2026, 18:41 IST)](https://www.cnbctv18.com/technology/china-ai-developers-publish-safety-tests-for-just-3-6-pc-of-model-releases-report-finds-20008702.htm)