Tencent’s Hy4 adds visual reasoning and computer control

Tencent launched Hy4 on Nov. 4, adding visual understanding and computer-use functions to its flagship AI model. The model can interpret detailed images and complex charts, reason with tool use, and operate both command-line tools and graphical interfaces.

The release extends the Hy4 Preview model introduced on Aug. 28. Both models have a one-million-token context window, but Tencent said it had substantially expanded pre-training and post-training for the full release to improve instruction following, long-context work, deep research and longer-running agent tasks.

Hy4 is designed for workflows that include web development, document summarisation and presentation, visual creative work, front-end engineering and game development. Tencent trained it with software such as Blender and CAD, as well as workflows that span multiple applications. It also worked with developers from Tencent Interactive Entertainment Group to curate game-development training data and evaluations.

For games, the company adapted Hy4 for Unreal Engine, Unity and Godot, including multiplayer development. Tencent reported improvements of 19.1% on its GameDev benchmark and 9.8% on GameLogic against Hy4 Preview, although the baseline scores were not released.

Tencent’s own tests also reported a 36.8% gain in multimodal agent-task completion over its previous vision-language model, and a 15.8% improvement in multimodal question answering. On 20 Honor of Kings configuration tasks, accuracy rose from 91.74% to 97.28%, reaching 99.32% after agent verification and self-repair. Its automated bug-analysis platform recorded root-cause-analysis accuracy rising from 59.5% to 73.8%.

The company said Hy4 reduced its hallucination rate from 15.76% in Hy4 Preview to 5.45%, while speculative errors fell from 23.1% to 5.6%. XenoSpectrum noted that Tencent’s official materials contained a separate wrong-answer measure, declining from 12.95% to 7.2%, and that the accounts addressed different targets. It also said Tencent had not disclosed test selection, test counts, metric definitions or aggregation methods for either account.

Hy4 is available through Tencent Cloud and OpenRouter, and in CodeBuddy, WorkBuddy, Yuanbao and ima. Tencent plans to open-source the model weights and publish a system card with safety evaluations and further technical details, as well as release a lower-cost Hy4 Flash variant with faster inference.