Tencent's Hy4 Lands With Vision, Computer Use, Game-Dev Gains

Tencent ships the full Hy4 model with computer use

Tencent Holdings today introduced Hy4, a new frontier large language model that adds computer-use capabilities for the first time, combining advanced visual understanding, reasoning and tool-calling for long-horizon agentic performance. The full release follows the Hy4 Preview from Aug.

28, which weighed in at 770 billion total parameters with 49 billion active parameters and a 1 million-token context window.

The company said it substantially scaled up both pre-training and post-training, reporting gains in instruction following, long-context understanding, agentic long-horizon tasks and deep research — while also fixing what it identified as excessive reasoning and over-verification in the preview's thinking processes.

Why this release stands out

The surprising part for working developers is how narrowly Hy4 targets real production workflows rather than generic chat.

According to unite.ai, targeted training covered Blender, CAD and long-horizon workflows spanning multiple applications, with vision connecting perception to planning and tool use across command-line and graphical interfaces.

The model is tuned for front-end work that many models fumble — complex UI interactions, business logic, state management and full-stack integration — and Tencent Interactive Entertainment Group developers curated game-development training data covering Unreal Engine, Unity and Godot, gameplay logic, asset pipelines and multiplayer development.

The model is available now through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao and WeCom's AI Assistant, plus third parties like OpenRouter, Hermes and OpenCode.

The evidence: benchmarks and safety checks

Tencent reports Hy4 improved by 19.1% on the GameDev bench and 9.8% on GameLogic over the preview, though the baseline numbers remain internal and weren't released.

According to unite.ai, the company also evaluated Hy4 with 213 finance professionals across 579 real-world tasks, where it ranked first among open-weight models with a score of 3.04, ahead of DeepSeek-V4.1-Flash-1017 at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68 each.

Hallucinations in internal search evaluations reportedly fell from 15.76% to 5.45%, and speculative errors dropped from 23.1% to 5.6%. Tencent Xuanwu Lab and Keen Security Lab conducted red-team audits, with cybersecurity training data limited to defensive proof-of-concept work.

Pricing holds, open weights and Flash are coming

Tencent acknowledges a remaining gap to top proprietary models on research-level mathematics, long-horizon coding and advanced visual question answering.

The company says it intends to release the model weights and system cards in coming days — the preview was open-sourced, unite.ai notes — and a faster, cheaper Hy4 Flash version is scheduled for release soon, with vLLM and SGLang ecosystem support planned alongside.

Sources

This one wasn’t built for you. The next one is.

Nobody wrote this. It came back from somebody’s Dots — standing orders they gave an agent, which go out, do the reading, and build what they find. Theirs are pointed at what they care about. Yours would be pointed at you.

Build your Dots. A market, a field, a question — anything you want to get good at.
They go. Research across the web and your accounts, checked against the sources.
They keep going. When something changes, your page updates and you hear about it.

Get ShareDot Browse more stories →