Tencent Hy4: 770B-parameter model can now see images and run your PC

What's new

Chinese internet giant Tencent unveiled the final version of its large language model Hy4 on November 4, 2026. It belongs to the Hunyuan family, which Tencent now simply calls "Hy." A text-only preview ("Hy4 Preview") came out back in August. According to Tencent, the final version received significantly more pre- and post-training and brings two capabilities for the first time: it can understand images and operate a computer.

Hy4 has 770 billion parameters in total and is built as a mixture of experts (MoE). That means the model consists of many specialized subnetworks, and only some of them kick in for each word. Per token, i.e. per word fragment, 49 billion parameters are active. The context window holds around one million tokens, which is how much text the model can take into account at once. That's enough for entire codebases or long work sessions.

You can already use Hy4 through Tencent Cloud TokenHub, the work and coding tools WorkBuddy and CodeBuddy, the chat assistant Yuanbao, the AI assistant "Dayuan" in WeCom (the business version of WeChat) and ima. An API is available through external platforms such as OpenRouter, Hermes and OpenCode. Tencent plans to publish the model weights and a system card (a technical document covering training and safety testing) "in the coming days." As of the available reports, they aren't yet freely downloadable.

New: seeing and clicking on its own

The preview could only handle text. According to Tencent, the final version can analyze images in detail, break down complex diagrams and solve math problems embedded in images.

The bigger deal is computer use. The model recognizes what's on the screen, controls the mouse and keyboard, and also works with the command line. That's meant to let it handle long workflows that span multiple applications. Tencent specifically trained Hy4 on professional software such as the 3D program Blender and CAD applications for technical drawing. The company talks about "real productivity scenarios" rather than pure benchmark chasing.

TechRepublic puts this in context: Western developers have so far relied mostly on closed systems such as Anthropic's Claude for automated UI control. An open model with this capability would be an alternative for companies that don't want to send sensitive screen content to third-party cloud providers. But the magazine also warns that operating graphical interfaces remains unpredictable. Small layout changes, rendering delays or unexpected pop-up windows frequently trip up such agents.

Numbers and benchmarks – all from Tencent itself

All figures come from Tencent's own or internal tests. An analysis on the platform note.com points this out too. There's no independent verification yet.

• Office work (WorkBuddy): An average of 83.18 points, a success rate of 96.15 percent and first place among Chinese models in 9 of 12 scenarios.
• Finance: A blind evaluation with 213 professionals, 153 of them from outside institutions and around 70 percent with more than five years of experience. They rated 579 real-world tasks without knowing which model was answering:

• On iRaB, an investment research benchmark from financial AI company Rabyte, Hy4 scores 12.03 percentage points higher than the preview.
• Hallucinations, i.e. made-up facts: according to internal measurements, the rate dropped from 12.95 to 7.2 percent. Errors based on unsupported assumptions fell from 23.1 to 5.6 percent. That covers cases like the model simply accepting a false premise in the question.
• Game development: Working with Tencent's gaming division IEG, Hy4 was adapted for Unreal Engine, Unity and Godot. It scores 19.1 percent better on the GameDev test and 9.8 percent better on the GameLogic test than the preview.
• Science: On TenureTrail, a test with 123 executable tasks across seven fields, the model reportedly stood out in areas such as mathematical constructions and physical modeling.

The author of the note.com post puts the 7.2 percent in concrete terms: the model gets roughly every 14th answer wrong. She therefore advises giving a screen-controlling agent only read-only tasks at first, then reversible steps such as saving drafts. Irreversible actions like sending, deleting or publishing should still be triggered by a human.

Self-criticism and safety

Tencent says Hy4 ranks "among the strongest open models for the first time." At the same time, the team admits the gap to the best closed models is still large, especially in advanced math, long software development tasks and demanding visual questions.

Because the weights are set to be released openly, Tencent is emphasizing safety. The model was trained on defensive security data, for example on patching vulnerabilities. Material on actively exploiting security flaws was deliberately excluded. Tencent's security labs Xuanwu Lab and Keen Security Lab tested areas including cryptographic reverse engineering (taking apart someone else's encryption), deobfuscating code and exploiting vulnerabilities. They found no significant offensive capabilities in the areas tested. On biosecurity, internal testing found the model has solid basic biology knowledge but only a limited ability to turn it into independent experimental protocols.

Tencent also describes a first step toward "recursive self-improvement." Using the tool Hyra 2.0, Hy4 optimized the training infrastructure for smaller models with the same architecture. This produced two open research models: HyCell for single-cell data and HyPDE for physical fields.

Pricing and availability

Pricing stays at preview levels. In China, the API costs 6 yuan per million input tokens, 18 yuan per million output tokens and 0.30 yuan for cached input. Tencent converts this to around $0.90, $2.70 and $0.045, while TechRepublic cites slightly different figures of $0.834 and $2.501. TokenHub offers a 50 percent discount through December 31, 2026, and WorkBuddy and CodeBuddy are free from November 4 to 15. A cheaper Hy4 Flash variant has been announced. QQ, Honor of Kings, Game for Peace, WeRead and Tencent News are also set to get Hy4. For the successor Hy5, Tencent is promising an even larger version.

What you need to try it

For now, the easiest way to try Hy4 is via OpenRouter or for free in WorkBuddy and CodeBuddy. The weights for running it locally haven't been released yet. And even once they are: at 770 billion parameters, even a heavily compressed 4-bit version would need roughly 400+ GB of memory. That takes multiple data center GPUs or very large workstations with lots of RAM, not a regular PC, Mac or Google Colab.

The bigger picture

It's a big leap: around nine months ago, the Hy3 series was still under 300 billion parameters. With Hy4, Tencent catches up with DeepSeek, Kimi and GLM and stakes out its own focus on computer use, something open models have barely offered so far. That could be especially interesting for companies that need to keep their data in-house. For individuals, the model is too big to run locally.

The major caveat remains that every performance figure comes from Tencent itself. The leads in the finance test are also narrow, and the office ratings are based on Tencent's own product. Whether computer use holds up outside of demos, coping with pop-ups, errors and changing interfaces, is something independent tests will have to show. And finally, the promise of "open" is only fulfilled once the weights are actually available for download under a clear license.