Tencent launches Hy4, its first AI model that can see images and operate a computer
Tencent's new flagship Hunyuan model adds image understanding and computer use for the first time, but its open weights and full safety report are still to come.
Chinese technology company Tencent has released Hy4, the full version of its flagship Hunyuan artificial intelligence model, adding two capabilities that its predecessors did not have: the ability to understand images and the ability to operate a computer through its graphical interface and command line.
Hy4 succeeds the Hy4 preview released in late August 2026, which was text-only. The full release, announced on 4 November, is built as a mixture-of-experts system with 770 billion total parameters, of which 49 billion are active for any given query, and a context window of about one million tokens, placing it in the same scale class as the largest frontier models currently available.
In practical terms, Tencent says the model can interpret fine-grained details in images, analyse complex charts and carry out mathematical visual reasoning, and can also work across professional software, including three-dimensional and computer-aided design tools. The company presents this as a step from a model that answers questions to one that can carry out longer, multi-step tasks inside real workflows.
Hy4 is available now as a hosted service rather than as downloadable software. It can be used through Tencent Cloud's TokenHub, through Tencent's own products including the Yuanbao chatbot, the WorkBuddy and CodeBuddy assistants, the ima application and the WeCom assistant Dayuan, and through third-party application programming interfaces such as OpenRouter. Tencent has kept the listed prices unchanged from the preview, at $0.834 per million input tokens and $2.501 per million output tokens, with a 50 per cent discount on TokenHub until the end of December and free use in WorkBuddy and CodeBuddy from 4 to 15 November.
One important qualification runs through the launch coverage. The model's open weights — the files that would let anyone outside Tencent download, run, fine-tune or independently inspect the model — have not yet been published. Tencent says it plans to release them, together with a comprehensive system card covering safety evaluations, in the coming days, with support from the vLLM and SGLang inference communities. Until that happens, Hy4 can only be used through Tencent's own services, and outside researchers cannot verify its internal behaviour directly.
Tencent's own claims for the model are ambitious. The company says Hy4 ranks among the strongest open-weight models for the first time, citing its performance across a table of benchmarks against rivals including DeepSeek, Kimi and GLM models, and reporting gains over the preview in instruction following, long-context understanding, long-horizon tasks and deep research. Those figures come from Tencent's own testing, and independent benchmarks and hands-on evaluations have yet to confirm them.
The launch also comes with a lighter version on the way. Tencent says a Hy4 Flash model, aimed at significantly better cost efficiency and faster inference, is scheduled for release soon, although it has given no specifications or date. The pattern — a large flagship first, a cheaper and faster variant afterwards — mirrors the release strategies of several rival laboratories.
Hy4 arrives in a fast-moving Chinese AI market. Moonshot AI recently released the weights of its Kimi K3.1 Pro model, DeepSeek remains a prominent open-weight rival, and Beijing is moving towards a comprehensive artificial intelligence law covering data, computing power and algorithms. In that context, Hy4 is as much a statement of Tencent's position in the domestic race as it is a product release, showing that its Hunyuan line can now handle images and computer operation alongside text.
What matters next is whether the promised open release materialises. If Tencent publishes the weights and the full safety report, independent researchers will be able to test the vision and computer-use claims, measure hallucination rates and compare the model fairly with its rivals. If the release slips, Hy4 will remain, for now, a powerful hosted service whose most interesting claims rest largely on its maker's word.
More in Technology & AI
Photo: Students working at computers in a school classroom (file image)
Technology & AIGoogle switches off AI tools for schoolchildren by default after criticism over classroom rollout
Google will switch off Gemini, Gemini Notebook, Gemini in Classroom and Search for school-managed student accounts until administrators explicitly turn them on, reversing its earlier decision to enable AI features for pupils by default.
Photo: Moonshot AI releases Kimi K3.1 Pro weights, a 2.8-trillion-parameter open-weight model
Technology & AIMoonshot AI releases Kimi K3.1 Pro weights, a 2.8-trillion-parameter open-weight model
China's Moonshot AI has published the full weights of Kimi K3.1 Pro on Hugging Face, giving researchers and developers free access to a 2.8-trillion-parameter multimodal model with a one-million-token context window.
Photo: Programming code on a screen (file image)
Technology & AIOpenAI to acquire Pierre Computer and fold its Code Storage system into Codex
OpenAI has announced plans to acquire Pierre Computer Company, the developer-tools startup whose Git-based Code Storage service and open-source Trees and Diffs libraries already power parts of its Codex coding app.