Tencent’s new Hy4 AI can now use a computer, and Tencent admits it still trails Claude and GPT

Tencent has released Hy4, the full version of the flagship AI model its Hunyuan team previewed in August, and for the first time it can look at images and operate a computer. Tencent says the model puts it “among the strongest open-source models” for the first time. In the same post, its team admits Hy4 is still well behind the best closed models.

Hy4 went live on Wednesday in Tencent’s own apps, including the Yuanbao chatbot, WorkBuddy, CodeBuddy, ima and the WeCom assistant Dayuan, and on Tencent Cloud’s TokenHub. Developers outside China can call it through OpenRouter, which lists a context window of about 1 million tokens and image input.

What’s new: vision and computer use

Compared with the Hy4 preview, Tencent says it scaled up both pre-training and post-training and improved instruction following, long-document understanding, long-running tasks and deep research. The model can now read charts, documents and photos, and in one demo it finds a roof fitting about 2 metres across in a 2.8-billion-pixel aerial photo.

Computer use is the bigger change. Hy4 can work through both the command line and on-screen apps, and Tencent says it trained it on professional software such as Blender and CAD tools and on long tasks across several apps. In Tencent’s own tests, its hallucination rate fell from 15.76% for the preview to 5.45%.

The team also makes bigger claims that haven’t been checked independently. It says Hy4, working with mathematicians at Peking University, found a counterexample to a published conjecture and verified the proof in Lean, and that the model is being prepared for “recursive self-improvement” by getting better at the work of building models.

Tencent admits a big gap to Claude and GPT

Unusually, the launch post has a section on Hy4’s limits. Despite being “one of the strongest open-source models”, the team writes, Hy4 still has “a significant gap” to the top closed models, especially on its hardest internal tests of frontier maths, long software projects, model development and visual questions.

Its own benchmark chart, which pits Hy4 against Claude Opus 5.5, GPT-6.1 Sol and three Chinese rivals, bears that out. On Terminal Bench 4.0, Hy4 scores 44.4, against 66.4 for Claude Opus 5.5 and 55.5 for GPT-6.1 Sol. On ProgramBench it gets 30.0 to Claude’s 56.5. It comes closest on OSWorld 2.1, a test of operating a computer, where it scores 78.9 to Claude’s 81.8, ahead of GPT-6.1 Sol’s 75.2. All the figures are Tencent’s, and several come from tests it built itself.

China’s open-weight race

Among open models, the picture is mixed. Hy4 leads DeepSeek’s V4.1 Flash and Z.ai’s GLM 5.3 on most of Tencent’s tests, but Moonshot’s Kimi K3.1 Pro beats it on several, scoring 66.8 on Terminal Bench 4.0. In a finance test with 213 experts across 579 real tasks, run with Tencent’s own fintech unit, Hy4 averaged 3.04 out of 4, against 2.97 for DeepSeek and 2.68 for both Kimi K3.1 Pro and GLM 5.3.

The weights aren’t out yet. Tencent says it “plans to” release them, with support from the vLLM and SGLang projects and the start-up Inferact, alongside a system card with fuller safety results, but gave no date. As of Wednesday morning, only the August preview, a mixture-of-experts model with 770 billion parameters (49 billion active), was on Tencent’s Hugging Face page, under the Apache 2.0 licence. Tencent hasn’t said which licence the full Hy4 will use. A cheaper, faster Hy4 flash is due “soon”.

Price and safety

Tencent kept the preview’s prices: 6 yuan (about $0.85) per million input tokens, 18 yuan (about $2.50) per million output tokens and 0.3 yuan for cached input. TokenHub calls are half price until December 31. OpenRouter currently lists Hy4 at $0.42 per million input tokens and $1.25 per million output tokens.

On safety, Tencent says it kept defensive security data in training, such as proof-of-concept generation and bug fixing, but removed exploit, reverse-engineering and penetration-testing material. Its Xuanwu and Keen security labs found no standout offensive cyber risk, it says, and Hy4 scored 30.4 out of 100 on its own biology-lab test, below what it calls the high-risk threshold. That is more than most Chinese labs publish at launch: an October study found only 9 of 857 models came with safety results, and Anthropic has warned about GLM 5.3’s cyber skills.

Why it matters

Hy4 makes Tencent a serious contender in China’s open-model race, with vision and computer use that most open rivals lack. But its own team’s admission is a useful reality check: the open models closing in on Claude and GPT on some charts still lag on the hardest work, and Hy4’s weights are a promise for now.

Sources: Tencent Hunyuan, “Hy4 正式发布” (in Chinese); Tencent, Hy4 announcement on Tencent News (in Chinese); OpenRouter, Tencent: Hy4; Hugging Face, tencent/Hy4-preview; Tencent, Hy4 preview announcement.