Tencent Hy4: A 770B AI That Sees Images and Operates Your Computer

Tencent released Hy4 on 4 November 2026 — a 770-billion-parameter mixture-of-experts model that does two things no Tencent model has done before: it can see images, and it can operate a computer. It is also, unusually, a release whose own benchmark chart shows it trailing the frontier. That honesty is what makes it worth a closer look.

What changed

Hy4 is the successor to the Hy4 preview from 28 August 2026. The preview was text-only; the full model adds vision and computer use, with a 1-million-token context window and 49 billion active parameters per token (per the preview's model card).

Vision covers fine-grained image processing, complex chart analysis, and mathematical visual reasoning. In one demo, the model found a roof fitting about two metres across inside a 2.8-billion-pixel aerial photo.

Computer use is the bigger shift. Hy4 can work through the command line and graphical interfaces — reading on-screen text, clicking buttons, and completing multi-step tasks across apps. Tencent says it trained the model on professional software such as Blender and CAD tools, and on long tasks spanning several applications.

The benchmarks, read honestly

Every number below comes from Tencent's own testing, where it re-ran rival models under consistent settings. None of it is an independent leaderboard.

The pattern is clear: on raw coding-agent work, Hy4 sits well behind Claude. On actually operating a computer, it is nearly level with Claude and ahead of GPT-6.1 Sol. Against other open models it is mixed — in a blind finance evaluation (213 professionals, 579 real tasks) Hy4 averaged 3.04/4, ahead of DeepSeek V4.1 Flash (2.97) and Kimi K3.1 Pro (2.68), but across Tencent's 12-benchmark table it ranks first on none of them.

Pricing is the real story

Tencent kept the preview's pricing: about $0.83 per million input tokens and $2.50 per million output tokens, with 50% off on TokenHub through 31 December 2026. OpenRouter lists it even lower, at $0.42/$1.25. For comparison, Gemini 4 Carbon costs $4/$20. A cheaper Hy4 Flash variant is due soon.

The catch

The full model's weights are not public yet. Tencent says they are coming "in the coming days" with a system card; only the August preview's weights are on Hugging Face under Apache 2.0. Until then, Hy4 is an API model — and a huge one: the preview's weights need roughly 924 GB even quantised, so self-hosting is a data-centre job, not a laptop one.

Hy4 is live now in Tencent's apps (TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom Dayuan) and via OpenRouter as tencent/hy4, with a free quota in WorkBuddy and CodeBuddy from 4–15 November.

Who it is for

If you need the best coding agent, Claude still wins — Tencent's own numbers say so. But if you need a cheap model that can look at a screen and operate software across a long task, Hy4 is one of the most interesting options at roughly a fifth of the frontier price. Judge it on the API, not the open-weights promise, until the weights actually land.

This is a condensed explainer. The full breakdown — the complete benchmark table, the Hy4 vs Kimi/DeepSeek/Claude comparison, pros and cons, and FAQ — is on NeurAITool: Tencent Hy4: 770B AI That Sees and Uses Your PC (2026).