Tencent unveils Hy4: an open model trained only on defensive cybersecurity
The Chinese giant showed a model with vision, computer use and stronger agentic skills. It excluded exploitation material from training and promises weights in the coming days.
What Hy4 can do
Hy4 adds visual understanding, including chart analysis and mathematical visual reasoning, and computer use across command-line and graphical interfaces. Tencent says it was trained on Blender and CAD, and targets frontend and backend engineering, game development in Unreal, Unity and Godot, financial analysis and scientific workflows.
On Tencent's WorkBuddy benchmark it averages 83.18 points with a 96.15% pass rate and finishes first in 9 of 12 mainland China scenarios. In a blind financial evaluation it scores 3.04, ahead of DeepSeek-V4.1-Flash (2.97), Kimi K3.1 Pro (2.68) and GLM-5.3 (2.68). The company reports its hallucination rate falling from 12.95% to 7.2% and unsupported assumptions from 23.1% to 5.6%.
Security by training data
Tencent took the opposite approach to OpenAI's GPT-6 Cyber. Instead of gating a dual-use model, it removed offensive material from the training data. The dataset focuses on proofs of concept and remediation, and Tencent says exploitation material was explicitly excluded. Its Xuanwu and Keen labs found no significant offensive-capability concerns in the tested areas.
That claim is self-reported and worth testing. The same week, Moonshot said it pruned cyberattack material from Kimi K3.1 Pro and still measured 81% on an internal exploitation test. Removing data reduces capability but does not necessarily remove it.
Price, availability and the open-source caveat
Hy4 is available through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom AI Assistant and ima, with APIs through OpenRouter, Hermes and OpenCode. Pricing is about $0.90 per million input tokens, $2.70 per million output tokens and $0.045 per million cached input tokens. WorkBuddy and CodeBuddy are free from 4 to 15 November, and TokenHub offers a 50% discount through the end of 2026.
The weights and system cards were promised for the coming days, so at launch Hy4 is open in name but not yet downloadable. A cheaper Hy4 Flash is planned. Until the weights actually appear, treat the open-source label as an intention.
Sources
More news
Moonshot publishes Kimi K3.1 Pro weights: 2.8 trillion parameters under a custom licence
The Chinese lab released full weights for a mixture-of-experts model with 104 billion active parameters, a 1M-token context and image input. It beats GPT-6 Astra on several coding benchmarks but trails Claude Opus 5.5 on the hardest tasks.
Read articleOpenAI releases GPT-6 Cyber, a model that found 81 Firefox vulnerabilities
A specialized model for authorized vulnerability research and exploitation is available only to vetted Daybreak Red customers. OpenAI rates its cybersecurity capability as Critical, the highest tier in its safety framework.
Read article