Hy4 Official Release, Built for Real Productivity

Building on the preview, the official Hy4 release expands pretraining and post-training scale, adds vision capabilities for the first time, and improves overall instruction following, long-document understanding, long-horizon tasks, and deep research.

For vision and computer operation, Hy4 handles fine-grained images, complex charts, and mathematical visual reasoning, supports combined command-line and GUI workflows, and is optimized for long-horizon tasks such as Blender and CAD.

In frontend and game development, Hy4 strengthens complex interactions, state management, and frontend-backend coordination, and supports engines such as UE, Unity, and Godot; GameDev and GameLogic benchmark scores improve by 19.1% and 9.8% versus the preview.

In finance and office scenarios, Hy4 is trained on real tasks and conducts expert crowdsourced testing with FiT; search optimization reduces hallucination rate from 15.76% to 5.45%, and it supports real-time collaborative editing in Word, Excel, and PPT.

For research and development, Hy4 ranks first among open-source models in internal TenureTrail evaluations, completing tasks such as cell reconstruction, mathematical conjecture correction, and training-system optimization; it is available on Tencent Cloud and OpenRouter, with plans to open-source the weights.