Tencent's Hy4 adds vision and computer use, and helped tune Tencent's own training pipeline

Tencent has shipped Hy4, the finished version of the model it previewed two months ago, with image understanding and computer use for the first time, Jonathan Kemper reports for The Decoder. Tencent says it ranks among the strongest open-weight models, though the weights themselves come later, on no fixed date. Until then, you can reach it through OpenRouter, Tencent Cloud and Tencent's own apps, Yuanbao among them.

What's new

• Vision: Hy4 is meant to catch fine detail in photos and dense diagrams. In one demo it found a roughly two-meter part on a warehouse roof in a 2.8-gigapixel aerial image, by writing code to crop sections.
• Computer use: it can drive software by itself. Tencent trained it on tools like Blender and CAD programs and on tasks that hop between apps.
• Game dev: with Tencent's own studios, it was tuned for Unreal, Unity and Godot. On Honor of Kings, the team says it handles configuration almost flawlessly and finds bug root causes far more often.
• Office work: it can edit Word, Excel and PowerPoint files alongside a person in real time. Tencent says a revamped search feature cut its hallucination rate during web research from almost 16% to about 5%.

The AI that trains AI

The most interesting claim: paired with Tencent's agent system Hyra 2.0, Hy4 optimized Tencent's own pipelines for training smaller models, raising throughput by about 20%. Next it trained its own take on Jev, a decision model built by the US startup TypeSafe AI, and that version beat the original on six of eight benchmarks. Tencent is releasing two biology and physics models made that way.

Benchmarks vs. vibes

Tencent says more than 200 finance professionals graded Hy4 on close to 600 real tasks and ranked it above GLM-5.3 Flash, Kimi K3.1 Pro and DeepSeek V4.1 Flash. Its own charts also show Claude Opus 5.5 ahead on all twelve benchmarks and GPT 6.1 Sol ahead on eight.

Independent numbers are less flattering. Artificial Analysis gives Hy4 35 on its Intelligence Index, above the median of 30, tied with Grok 4.7 and second among Chinese labs behind Kimi K3.1 Pro at 46. It roughly matches Anthropic's Haiku 5.5, which costs nearly half as much per task. Without web search, it answers about half of factual questions correctly, and when it doesn't know, it invents an answer 89% of the time.

Pricing stays at the preview's roughly $0.83 per million input tokens and $2.50 per million output. Tencent admits it lags the best commercial models on hard math, long coding projects and complex image analysis, and a cheaper Hy4 flash is next.

If you plan to use it, keep search turned on.

Scoreboard update: AI +1. Tencent's new model helps speed up the pipeline that trains Tencent's next models.

Stay undead.