Latest
AI Affairs, home

Sunday 22 November 2026

Technology

Tencent publishes Hy4 safety report with assessments from outside groups

Alice’s evaluation assessed Hy4’s cyber-hacking abilities as limited, according to the safety report described by the South China Morning Post.

Tencent headquarters campus buildings at dusk, framed by trees and an empty plaza
Photo: Lhzss8, CC BY-SA 4.0, via Wikimedia Commons (cropped)

Shenzhen-based Tencent published an 82-page safety report on 20 November for Hy4, drawing on outside assessments of the model’s hacking and behavioural risks ahead of its open-weight release, the South China Morning Post reported on 22 November. The newspaper described it as the first disclosed pre-release evaluation of a model from a leading Chinese technology company by independent third parties.

Key points

  • Alice assessed Hy4’s hacking capability as limited; Fangcun AI found low rates of deception and loss of control.
  • On Tencent’s own biological-experiment test, Hy4 scored three times higher than Hy4 preview.
  • Tencent released downloadable Hy4 weights on 20 November, allowing developers to run the model themselves.

Hy4 weights open up deployment

Following Hy4’s initial launch on 4 November, Tencent released its model weights for developers to download and deploy on 20 November, TechNode reported. AI Affairs covered the earlier Hy4 launch and its visual reasoning.

Hy4’s mixture-of-experts language backbone has 770 billion parameters, but activates 49 billion for each token, or piece of text, it processes. Rather than putting every token through the entire model, this design routes it through a selection of components.

Office tasks are among the workflows Hy4 supports, alongside frontend and game development, TechNode reported.

Alice and Fangcun AI assess Hy4

Chinese start-up Fangcun AI and Israeli-American firm Alice evaluated the open-weight model. According to the safety report as described by the South China Morning Post, Alice found that Hy4’s cyber-defence capability had improved over Hy4 preview, but assessed its hacking capability as “limited”. Fangcun found low rates of deception and “loss of control”, meaning the model largely followed its engineers’ instructions.

Tencent also thanked Singaporean AI safety organisations, including Neo Research, for contributions to the evaluations in a subsequent post on X.

Tencent’s Hy4 biological test records a higher score

Tencent said it restricted high-risk biological training data and trained Hy4 to refuse dangerous requests, including those concerning pathogen modification, the South China Morning Post reported. Its chief AI scientist, Vinces Yao Shunyu, had warned of biological threats from AI at a Hong Kong science forum earlier in November.

On Tencent’s own test of experimental design for wet-lab biology, Hy4 scored three times higher than Hy4 preview. Tencent said it found no evidence that Hy4 significantly lowered the barrier to biological harm because it had difficulty turning lab plans into dependable sequences of steps and could not supply the hands-on expertise for novices to carry out complex laboratory work.

Sources

Topics: Foundation models, Open source, Safety

What should we cover? Tell us

The weekly briefing

The AI Affairs weekly briefing: the week's 10 most important AI stories in business, government and markets, every Monday. We'll email you a link to confirm. Every email has a one-click unsubscribe. Privacy policy