On November 4, 2026, Tencent released Hy4, the latest generation of its Hunyuan large language model. It is a 770-billion-parameter model built on a mixture-of-experts architecture, and Tencent says it now ranks among the world’s leading open-source models. The launch is notable for two reasons at once: a major capability jump, and a deliberate safety choice. In a week when AI-driven cyberattacks are dominating security headlines, Tencent trained Hy4 with its cybersecurity data deliberately limited to defensive material.
What Tencent launched on November 4
Hy4 expands on its preview version with larger pre-training and post-training, and it adds visual understanding for the first time. It can read fine detail in images, analyse complex charts and work through visual maths problems with reasoning. Tencent is also releasing it across its own products, including Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao and WeCom’s AI assistant, with API access through third-party platforms such as OpenRouter. Model weights and a full system card are promised in the coming days, so for now this is best understood as a hosted launch on the way to a full open release. A lighter Hy4 Flash version focused on cost efficiency is also coming soon. Read Tencent’s official Hy4 announcement for the full details.
Defensive by design: what was filtered out
This is the part of the story that sets Hy4 apart. According to Reuters and Tencent, Hy4’s cybersecurity training data was limited to defensive content such as proof-of-concept generation and vulnerability remediation. Material covering exploitation, reverse engineering and penetration testing was filtered out, with the aim of limiting the model’s potential offensive capabilities. Tencent says experts from its Xuanwu Lab and Keen Security Lab audited the model across areas including cryptographic reverse engineering, binary deobfuscation and vulnerability exploitation, and found no significant concerns in the areas evaluated. Tencent’s own statement puts it plainly: “We recognize the potential safety risks posed by advanced cyber and biological capabilities, and we are committed to a more responsible approach to model training, open-sourcing, and release.” The same announcement notes a biosafety limit too: Hy4 has strong foundational biological knowledge, but its open-ended experimental design ability is more limited than frontier models. That honesty matters, because a safety filter is a design choice, not a guarantee, and independent testing will be the real test once weights are public. See the Reuters report on Hy4’s limited cybersecurity training for more on this approach.
Vision and Computer Use: the capability leap
Hy4’s headline new feature is Computer Use, which lets it work through both command-line interfaces and graphical user interfaces to carry out long, multi-step tasks. Tencent has trained it on professional software such as Blender and CAD, and says it can plan and execute sophisticated work across applications. That puts it squarely in the agentic AI trend, where models are judged less on answering a single question and more on completing a whole job. For everyday users, the more relatable change may be visual understanding: point the model at a chart, a photo or a document image, and it can reason about what it sees rather than only reading text.
Real-world tests: finance, office, coding and science
Tencent is pitching Hy4 on practical work, not just benchmarks. In its WorkBuddy office evaluation, Hy4 scored an average of 83.18 with a 96.15 percent pass rate, ranking first in nine of twelve scenarios in mainland China. In finance, a blind evaluation with 213 financial professionals across 579 real tasks put Hy4 at 3.04, ahead of DeepSeek-V4.1-Flash-1017 at 2.97, Kimi K3.1 Pro at 2.68 and GLM-5.3 at 2.68. Tencent also reports its hallucination rate fell from 12.95 percent in the Hy4 Preview to 7.2 percent, with errors from unsupported assumptions falling from 23.1 percent to 5.6 percent. In game and front-end development, GameDev scores improved 19.1 percent and GameLogic 9.8 percent over the preview, with support for Unreal Engine, Unity and Godot. In science, a 123-task TenureTrail evaluation covered maths, physical modelling, earth sciences and systems engineering. These are vendor-reported figures, so treat them as a launch claim to be checked against independent benchmarks, not as settled fact. ChainCatcher’s Hy4 evaluation coverage provides further detail on these results.
Pricing, availability and the open-source catch
Despite the capability jump, Tencent is holding Hy4’s price at about US$0.90 per million input tokens, US$2.70 per million output tokens and US$0.045 per million cached input tokens. From November 4 to 15, WorkBuddy and CodeBuddy are free to try, and Hy4 on Tencent Cloud TokenHub is discounted 50 percent through December 31, 2026. The important caveat is timing: until the weights and system card actually appear, researchers cannot fully verify the safety claims or run the model themselves. Tencent itself acknowledges a gap remains between Hy4 and the top closed models. The launch is therefore a promise in two parts, capability now and verifiability shortly after, and the second part is the one to watch.
Why Hy4 matters — and what to watch next
Hy4 matters because it joins two debates that are usually kept apart: how powerful open models are becoming, and how much safety responsibility should sit in training data itself. Moonshot AI limited its newest model’s cyber-offensive capabilities last month, and Hy4 extends that pattern among Chinese labs, alongside similar caution from US companies. It also lands as China moves toward broader AI legislation, covered in our report on China’s full-lifecycle AI law, and as enterprises weigh agent controls such as CrowdStrike’s AI kill switch. The launch also fits a wider pattern of national AI efforts, including India’s homegrown EKA-1 and Dhīr models, and it arrives as enterprise AI spending hit $177 billion in 2026. Watch three things next: the actual release of weights and the system card, independent safety evaluations once researchers can test the model, and whether defensive-only training becomes a standard other open-model builders adopt or a one-off experiment. If it holds up under outside scrutiny, Hy4 could become a template for open models that are powerful without being easy to weaponise.
