Tencent launches the full Hunyuan Hy4 with vision and computer-use capabilities

Tencent has released the full version of Hunyuan Hy4, its latest-generation large language model, roughly two months after an open-weight preview appeared in late August. The headline additions are visual understanding, a first for the Hunyuan line, and computer-use capability, letting the model operate software through graphical interfaces and command lines. Caixin reported the launch on 4 November and described it as an attempt to narrow the gap with leading AI developers.

Compared with the preview, Tencent says Hy4 improves instruction following, long-document understanding and long-horizon agent behaviour, and targets complex real-world tasks including office collaboration, financial analysis, front-end development and game development. Combined with reasoning and tool calling, the vision capability is pitched at fine-grained image processing, chart analysis and mathematics problems that include diagrams.

Caixin's digest puts the model at 770 billion parameters, up from under 300 billion for the previous generation, and reports a hallucination rate of 7.2 per cent, down from 12.95 per cent, with "guessing errors" falling from 23.1 per cent to 5.6 per cent. Input pricing is listed at 6 yuan per million tokens. These figures come from Tencent's own materials and Caixin's automated summary rather than independent evaluation.

Hy4 is available through Tencent Cloud TokenHub, the WorkBuddy and CodeBuddy developer tools, the Yuanbao consumer assistant, ima and the WeCom intelligent assistant Dayuan. Reports differ on whether the weights of the full release will be opened as the preview's were; Tencent has said open-sourcing will follow but has not given a date. The launch coincided with Chinese models holding the top four places in OpenRouter's weekly token rankings.

Why it matters

Hy4 is another data point in a pattern: Chinese labs are shipping frontier-adjacent, multimodal, agent-capable models at a cadence and price that Western vendors struggle to match. Even for buyers who will never deploy a Chinese model, the competitive pressure on pricing and feature parity is real.

Indicops view

For Indian organisations the direct relevance is limited by data-residency and supply-chain concerns, but the capability trend is not. Vision-plus-computer-use is now table stakes for any serious agent platform, and procurement specifications written around text-only models are already dated. We advise clients to specify multimodal document handling and GUI automation as evaluation criteria now, regardless of which vendor ultimately wins the contract.