Tencent's Hy4 Lands With Vision, Computer Use and Top Benchmarks

Tencent pulled the wraps off Hy4, its next-generation large language model, and says it now sits among the world's top open-source AI systems. The company plans to publish model weights and system cards within days. It's the first model in the Hy line to understand visuals, and the first that can actually operate a computer, working through command-line tools and graphical interfaces alike.

Tencent trained Hy4 on professional software including Blender and CAD, sharpening its ability to plan and finish long, multi-step jobs. The accuracy push shows up in the numbers: internal tests cut the hallucination rate to 7.2%, down from 12.95% in the earlier Hy4 Preview. Errors traced to unsupported assumptions dropped to 5.6%, down from 23.1%.

In a blind financial evaluation run with Tencent Financial Technology, 213 finance professionals judged 579 real-world tasks. Hy4 came out on top with a score of 3.04, ahead of DeepSeek-V4.1-Flash-1017 at 2.97, Kimi K3.1 Pro at 2.68 and GLM-5.3 at 2.68. Roughly 70% of the 153 outside experts had more than five years in the industry. Hy4 also posted a 12.03 percentage-point gain over Hy4 Preview on the iRaB benchmark from financial AI firm Rabyte.

On Tencent's WorkBuddy, Hy4 scored an average of 83.18 with a 96.15% pass rate, taking first place in nine of twelve scenarios across mainland China. Wired into WeCom through WorkBuddy, it tracks chats, emails, documents, calendars and meetings so users juggling multiple projects can spot overdue work and looming deadlines. A team collaboration layer arrives with WorkBuddy's upcoming "Projects" feature.

Developers get attention too. Hy4 improved 19.1% on GameDev and 9.8% on GameLogic benchmarks versus the preview, with deep hooks into Unreal Engine, Unity and Godot for complex game logic and multiplayer builds. In science, it was validated on 123 executable tasks across seven disciplines using TenureTrail. Tencent Hy says it's moving toward recursive self-improvement with its Hyra 2.0 harness, and it open-sourced HyCell for single-cell representations and HyPDE for physics fields.

On safety, Tencent says Hy4 trains only on defensive security data, prioritising proof-of-concept work and vulnerability fixes while excluding offensive exploitation material. Experts from Xuanwu Lab and Keen Security Lab assessed reverse engineering, binary deobfuscation and vulnerability exploitation, and flagged no significant concerns. Internal reviewers found solid foundational biology knowledge but weaker open-ended experimental design. Hy4 is live on Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom's AI Assistant (Dayuan) and ima, with its API on OpenRouter, Hermes and OpenCode. More Tencent integrations, including QQ, Honor of Kings, Game of Peace and WeRead, are in the works.