Hy4 leaves preview: Tencent adds vision and computer use, but the weights are only announced

On 4 November Tencent announced the official version of Hy4, ending the preview phase that opened on 28 August. According to the company's page, the model brings two capabilities for the first time. The first is visual understanding: fine-grained image reading, analysis of complex charts and mathematical reasoning over visual material. The second is computer use, meaning it can act both from the command line and through graphical interfaces. Tencent writes that training included professional software such as Blender and CAD programs, and that the model works with Unreal Engine, Unity and Godot. The company lists Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, the WeCom AI assistant and ima as channels, plus API access through third-party platforms such as OpenRouter, Hermes and OpenCode. Integrations into QQ, Honor of Kings, Game of Peace, WeRead and Tencent News are under way.

Official pricing is unchanged from the preview: about $0.90 per million input tokens, $2.70 per million output tokens and $0.045 per million cached input tokens. TokenHub offers a 50% discount until 31 December 2026, while use on WorkBuddy and CodeBuddy is free from 4 to 15 November. TechNode, by contrast, reports $0.834 and $2.501, figures that do not match Tencent's.

On the weights, the company says it intends to publish the model weights and system cards in the coming days, with particular attention to cybersecurity and biosecurity, but gives no date. When I checked, Hugging Face only had the preview collection, so the licence and system card of the official version cannot be verified.

The model's size is also uncertain. The official page states neither the parameter count nor the context window. SiliconANGLE reports 770 billion parameters, 49 billion of them active, and a 1-million-token context. TechNode gives 770 billion, while the Hugging Face card for the preview lists 780 billion.

All the results Tencent publishes come from internal evaluations or from evaluators chosen by the company, and none include comparisons with closed models. In the WorkBuddy evaluation, Hy4 scores an average of 83.18 with a 96.15% pass rate, and ranks first in 9 of 12 scenarios in the Chinese market. In a blind finance test with 213 professionals and 579 tasks, Hy4 scores 3.04, against 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 for both Kimi K3.1 Pro and GLM-5.3. According to Tencent, the hallucination rate drops from 12.95% in the preview to 7.2%, and errors caused by unsupported assumptions fall from 23.1% to 5.6%. The company also claims improvements of 19.1% on GameDev and 9.8% on GameLogic over the preview, but SiliconANGLE notes that the absolute starting values were not published. On cybersecurity, the model was assessed by Tencent Xuanwu Lab and Keen Security Lab, and Tencent says it excluded offensive material related to vulnerability exploitation from training. In the biosecurity section, the company acknowledges that Hy4's capabilities in open-ended experimental design are more limited than those of frontier models. The announcement is rounded out by two open-source scientific models built with Hy4 and Hyra 2.0: HyCell, for single-cell representations, and HyPDE, for physical fields. Then there is Hy4 Flash, a cheaper version. Its status is ambiguous on the same page: the text says it is “coming soon”, while a heading labels it “Recently Released”.

Tencent places Hy4 among the world's leading open-source models, but for the official version that openness exists, for now, only as a written commitment in the announcement: what sits on Hugging Face is the preview collection. Anyone who wants to try it goes through the company's APIs and products, and the numbers available to judge it are the ones Tencent chose to measure and publish. I find the biosecurity sentence more instructive: it is one place where the announcement points to a limit of the model compared with frontier models, in the company's own words. Once the weights and system cards arrive, the rest can be checked.

— Olya