Tencent Hy4 arrives with eyes and hands, but its weights stay behind

A full release, not a preview

Tencent released the full version of Hy4 on 4 November 2026, following the preview released in August. The company says it substantially scaled up pre-training and post-training, and describes Hy4 as joining the strongest open-source models for the first time.

The model is available now through Tencent Cloud and OpenRouter and inside CodeBuddy, WorkBuddy, Yuanbao and ima. Listed API prices are unchanged from the preview: 0.834 dollars per million input tokens, 2.501 dollars per million output tokens and 0.042 dollars per million cached input tokens.

[1] [2]

What it can do now

Hy4 adds vision perception for the first time: it can read fine-grained image details and interpret complex charts, combining perception, reasoning and tool use.

It also supports computer use for the first time, working across command-line and graphical interfaces. Targeted training covered professional software such as Blender and CAD and long tasks spanning several applications.

In Tencent's own evaluations, Hy4 ranked first among open-weight models in a finance test scored 3.04 out of 4, ahead of DeepSeek-V4.1-Flash at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68 each. On computer-use and terminal tests its reported scores trail the leading closed models: 78.9 on OSWorld 2.1 against 81.8 for Claude Opus 5.5, and 44.4 on Terminal-Bench 4.0 against 66.4 for Claude Opus 5.5 and 66.8 for Kimi K3.1 Pro. Tencent also reports an internal search hallucination rate falling from 15.76 percent for the preview to 5.45 percent. These are vendor-run figures, not independent results.

[1] [2] [3]

What is not open yet

The full model's weights are not yet downloadable. Tencent says it plans to open-source them, with ecosystem support from the vLLM and SGLang communities, and will publish a comprehensive system card with safety evaluations when the weights arrive. No release date for the weights was given in the material reviewed, and the licence for the full version was not stated.

Tencent itself acknowledges a notable gap to top proprietary models, especially on research-level mathematics, long-horizon coding, model self-improvement and advanced visual question answering. A separate Hy4 Flash version is planned for the near future, promising better cost efficiency and faster inference, but no date was given.

[1] [2] [3]

How to read this launch

The honest summary is hosted capability now and open deployment later: teams can call Hy4 through Tencent's products and APIs today, but they cannot yet run the full model on their own servers, and the August preview's weights are not the new model.

Treat the strongest-open-model claim as a vendor claim until three things arrive: the weights and their licence, the promised system card, and independent evaluations. Until then, the launch is best judged on the hosted offering and its price, not on a ranking.

[1] [2] [3]