Skip to content
sobolewski.proAI today, every eveningSubscribe

Once a day. Unsubscribe with one click. Privacy policy

Tencent Hy4: Tencent's first model with computer use

Tencent Hy4 understands images, supports computer use and costs about $0.90 per million input tokens, with open weights on the way. Here is what it can do.

3 min read

This article in other languages:العربيةDeutschEspañolFrançaisItaliano日本語Polski简体中文

In this article
  1. What’s new in Tencent Hy4?
  2. Why is everyone talking about it?
  3. What to watch out for?
  4. How to try Tencent Hy4?
  5. FAQ

On November 4, Tencent announced Tencent Hy4, the full version of the model whose Hy4 Preview came out on August 28. It is Tencent’s first model that understands images and supports computer use, and the weights and system cards are due to be published “in the coming days”.

What’s new in Tencent Hy4?

According to Tencent’s announcement, Hy4 is the first model in the series to get image understanding and computer use, meaning it can operate a computer both from the command line and through a graphical interface. Interestingly, the model also went through dedicated training on Blender and CAD software.

Beyond that, Tencent improved instruction following, long-context understanding and long agentic tasks. The Preview version is an MoE (mixture-of-experts) model with 770B parameters, 49B of them active, and both versions have a 1M-token context window (as reported by SiliconANGLE).

The most interesting numbers from the announcement:

  1. In internal search tests, the hallucination rate fell from 12.95% (Hy4 Preview) to 7.2%, and errors caused by unfounded assumptions fell from 23.1% to 5.6%.
  2. On the GameDev and GameLogic benchmarks, the score rose by 19.1% and 9.8% over the Preview (without the baseline scores).
  3. In WorkBuddy, Tencent’s agentic app, the model reached an average score of 83.18 and a 96.15% task pass rate.

These are all Tencent’s own measurements, so take them with a grain of salt.

A curiosity from the lab is Hyra 2.0, an autonomous research agent. With its help, Hy4 optimized the pre-training infrastructure for smaller models with the same architecture, and Tencent released two scientific models built this way: HyCell (single-cell representations) and HyPDE (physical fields).

Why is everyone talking about it?

First, open weights at large scale. Another Chinese lab is promising the weights of a model with 1M tokens of context and agentic features, less than two weeks after Kimi K3.1 Pro, whose weights Moonshot has just pushed to Hugging Face.

Second, the price. According to Tencent, Hy4 costs about $0.90 per million input tokens, $2.70 per million output tokens and $0.045 per million cached tokens, the same as the Preview. On top of that, Tencent Cloud TokenHub offers a 50% discount until December 31, 2026.

Third, a decision on cybersecurity. Tencent writes that offensive material related to exploiting vulnerabilities was excluded from the training data, and tests by Xuanwu Lab and Keen Security Lab found no “significant risks” in the areas examined. In the Hacker News discussion opinions are split: some commenters consider it sensible, others say they will stick with other models for that kind of work. One commenter also noted that the model can be very slow (in their test it made up to 64 requests for a single question).

What to watch out for?

First, the weights. As of launch day they are only a promise, and Tencent has not yet given a license for the full Hy4.

Second, biology. Tencent itself admits that the model is more limited than frontier models at designing open-ended biology experiments.

Third, the Flash version. The cheaper Hy4 Flash is due “soon”, with no specific date or specs.

How to try Tencent Hy4?

The model already runs in several places:

  1. Tencent apps: Yuanbao, ima, the AI assistant in WeCom (Dayuan), plus WorkBuddy and CodeBuddy, which are free from November 4 to 15.
  2. Tencent Cloud TokenHub, with the 50% discount mentioned above until the end of the year.
  3. API via third-party platforms: OpenRouter, Hermes and OpenCode, so we can plug Hy4 into the agent harness we already use.

Once the weights are out, Hy4 will join the models you can run yourself, although at this scale it is a job for a server room rather than a laptop (more about hardware in the guide how to run AI locally).

FAQ

Does Tencent Hy4 have open weights?

Not yet. Tencent promises to publish the weights and system cards in the coming days, but it has not given a license.

How much does Tencent Hy4 cost?

According to Tencent, about $0.90 per million input tokens and $2.70 per million output tokens. Tencent Cloud TokenHub offers a 50% discount until December 31, 2026, and WorkBuddy and CodeBuddy are free until November 15.

What can Tencent Hy4 do?

It understands images, operates a computer from the command line and through a graphical interface, went through dedicated training on Blender and CAD, and has a 1M-token context window.

  • AI
  • open-weight
  • AI agents
  • API

AI today, every evening

The day's biggest AI news and new posts from the blog, every evening. Short, with sources.

Once a day. Unsubscribe with one click. Privacy policy