Tencent Launches 770B Hy4 Model with Vision and Computer Use
Tencent just dropped Hy4, a 770B parameter open-source model that can see your screen, click buttons, and run command lines.
Models · Source: ChainCatcher
What happened
Tencent officially released its new generation large language model, Hy4, on November 4. It is a massive 770 billion parameter model built on a mixed expert architecture. The model weights will be open-sourced soon, but developers can already access it via Vercel AI Gateway. Tencent is also preparing for Recursive Self-Improvement, using the model to automatically train smaller infrastructure models.
The biggest upgrade is vision combined with Computer Use. Hy4 is the first in the series that can read complex charts, solve visual math problems, and recognize fine image details. It uses this vision to operate graphical user interfaces and command lines directly. Tencent specifically trained it to use professional software like Blender and CAD, enabling cross-application collaboration for long-range tasks.
The model is already proving itself in real workflows and specialized industries. It scored 3.04 in a blind financial test with 213 experts, beating models from Kimi and DeepSeek. It also integrates directly with WeChat Work to manage emails, documents, and schedules across multiple projects. For scientific research, Tencent evaluated it across seven disciplines and open-sourced foundation models for single-cell characterization and physical fields.
Key facts
- 770B — Total parameter scale of the Hy4 model
- 1M — Token context window size available on Vercel
- 3.04 — Hy4 score in financial crowd-testing
- 96.15% — Task pass rate in the WorkBuddy office scenario
- 19.1% — Improvement in GameDev evaluation scores
Why it matters
AI is moving from a chatbot that answers questions to an agent that actually does the work. With Computer Use and a massive one million token context window, you can build systems that navigate legacy software, read visual data, and execute multi-step tasks autonomously. You no longer need complex API integrations for every single tool in your stack. The model can simply look at the screen, understand the interface, and click the right buttons to get the job done. This drastically reduces the friction of building AI agents for traditional desktop environments.
The open-source landscape is getting highly competitive for agentic workflows. Tencent is aggressively targeting specialized verticals like front-end development, multiplayer online games, and financial analysis. They are optimizing for complex interactions and state management in engines like Unreal Engine and Unity. When a 770 billion parameter model with these capabilities goes open-source, it forces proprietary model providers to lower their prices. It also pushes the baseline for what free models can achieve, giving founders a powerful new engine for building autonomous software operators.
For builders
Cheaper inference on Vercel AI Gateway
Hy4 is available right now on Vercel AI Gateway under the model name tencent/hy4. Vercel is offering it at 50 percent off through December 31, 2026. This significantly lowers the barrier for founders who want to test long-context tasks across large codebases without paying premium provider fees.
Automate game dev and legacy interfaces
Hy4 features deep adaptation for major game engines like Unreal Engine, Unity, and Godot. Combined with Computer Use, engineers can build agents that manipulate game assets or operate legacy software interfaces directly. You save weeks of custom API wiring because the AI interacts visually, leaving competitors who rely on rigid integrations behind.
Lower hallucination risk in fintech apps
Tencent optimized Hy4 for multi-source information integration and factual consistency verification. Internal testing showed hallucination rates dropped from 12.95 percent to 7.2 percent compared to the preview version. Fintech builders get a much more reliable base model for analyzing complex market data, reducing the risk of costly trading errors.
My take
I have been saying AI needs to move from just thinking to actually doing. Tencent just proved it by putting Computer Use into a massive 770 billion parameter open-source model. When AI can take control of the mouse and keyboard to operate software like Blender, agentic automation becomes a cheap commodity for every builder.
Original reporting: ChainCatcher. This is my rewrite and opinion.