Tencent released the official version of Hy4, its latest Hunyuan large language model, on Wednesday, and said it trained the model with limited cybersecurity data to reduce the chance it could help with cyberattacks. Tencent said it filtered out training material on exploit development, reverse engineering and penetration testing, keeping the focus on defensive security content.
Tencent framed the data choice as a safety measure. "We recognize the potential safety risks posed by advanced cyber and biological capabilities," the company said in a statement. It said it is pursuing a more responsible approach to open-sourcing, with attention on network security and biosecurity risks, and plans to publish a model system card with more comprehensive safety evaluations alongside the weights.
Those weights are not out yet. Tencent describes Hy4 as an open-source model and says the weights will be released soon, without giving a date. It has worked with the vLLM and SGLang communities and Inferact Inc. to prepare inference support.
Key facts
- 01Tencent released the official Hy4 model on November 4, 2026, following the Hy4 preview on August 28.
- 02Tencent says it filtered exploit development, reverse engineering and penetration testing material out of Hy4's training data.
- 03Internal hallucination rate fell from 12.95% in the preview to 7.2%, according to Tencent.
- 04API pricing is unchanged at 6 yuan per million input tokens and 18 yuan per million output tokens.
- 05Tencent says the weights will be open-sourced soon, alongside a system card with safety evaluations.
Hy4 follows the Hy4 preview Tencent released on August 28, which had 770 billion total parameters, 49 billion active parameters and a context window of more than 1 million tokens. Tencent's previous Hy3 series was under 300 billion parameters. Compared with the preview, Tencent says it expanded both pretraining and post-training, with gains in instruction following, long-text understanding, long-horizon tasks and deep research.
The official release adds two capabilities the preview lacked. Hy4 is the first Hunyuan model with visual understanding, combining perception, reasoning and tool calling for image processing, chart analysis and visual math. It is also the first to support computer use through both the command line and graphical interfaces, and Tencent ran specialized training on professional software such as Blender and CAD.
Tencent's self-reported numbers center on work tasks. On its WorkBuddy office evaluation, Hy4 averaged 83.18 with a 96.15% task pass rate and placed first among domestic models in 9 of 12 scenarios. In a finance test where 213 experts blind-rated 579 real tasks, it scored 3.04, which Tencent says is the top result among open-source models. Its internal hallucination rate fell from 12.95% in the preview to 7.2%, and errors from guessing dropped from 23.1% to 5.6%.
Tencent is direct about where Hy4 stands. On code editing, automation and computer-use benchmarks, it says Hy4 beats DeepSeek-V4.1-Flash but trails GPT-6.1 Sol and Claude Opus 5.5, and it acknowledges a remaining gap with top closed-source models on its hardest internal tests in frontier math, long-horizon development and visual question answering.
Pricing holds at the preview's level: 6 yuan per million input tokens, 18 yuan per million output tokens and 0.3 yuan per million cache-hit tokens. WorkBuddy and CodeBuddy offer free use from November 4 to 15, and Tencent Cloud TokenHub charges half the list price through December 31.
Hy4 is live on Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, ima and a WeChat Work assistant, and through OpenRouter for API access. Tencent is integrating it into QQ, Honor of Kings, Peacekeeper Elite, WeChat Reading and Tencent News, and says a cheaper, faster Hy4 flash variant is coming soon.
The release follows a similar move from Moonshot AI, which limited the cyber-offensive capabilities of its newest model last month.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
