Tencent releases Hy4, its first model that can read images and operate a computer

What is new in Hy4

Hy4 is the full release of the model Tencent previewed on August 28. That preview was a mixture-of-expertsa design in which only part of the model, a few "experts", handles each request, so it costs less to run than its full size suggests model with 770 billion parameters, 49 billion of them active, and a context window of one million tokens. The November announcement gives no new figures, so it is not clear whether the size changed.

Two abilities are new. Hy4 can understand images, which Tencent says covers detailed image processing, complex charts and visual math problems. And it can use a computer: it works with both command-line interfaces and graphical interfaces. Tencent says it trained the model on professional software such as Blender and CAD tools so it can plan and carry out longer tasks.

What Tencent claims

All of the numbers in the announcement come from Tencent's own tests or tests run with partners, not from independent evaluations:

• On WorkBuddy, Tencent's office assistant, Hy4 averaged a score of 83.18 with a 96.15% pass rate and ranked first in nine of twelve scenarios in mainland China.
• In a blind test of 579 financial tasks judged by 213 finance professionals, run with Tencent Financial Technology, Hy4 scored 3.04, ahead of DeepSeek-V4.1-Flash-1017 (2.97), Kimi K3.1 Pro (2.68) and GLM-5.3 (2.68).
• In internal tests, its hallucinationan answer the model states as fact but that is made up or wrong rate fell from 12.95% in the preview to 7.2%.
• Compared with the preview, it improved 19.1% on a game development benchmark and 9.8% on a game logic benchmark.

The comparisons in the announcement are with other open models. Tencent does not compare Hy4 with the closed models from OpenAI, Google or Anthropic.

Safety claims

Tencent says it kept material about offensive exploitation out of the training data, and that experts from its Xuanwu Lab and Keen Security Lab found "no significant concerns" about the model's offensive cyber abilities in the areas they tested. On biology, it says Hy4 has strong basic knowledge but is weaker than frontier models at designing open-ended experiments. These are the company's own assessments. The system card that should explain them in detail has not been published yet.

Where to use it and what it costs

Hy4 is available in Tencent's own products, including Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, the Yuanbao assistant and ima, and through third-party platforms such as OpenRouter. The price is about $0.90 per million input tokens, $2.70 per million output tokens and $0.045 per million cached input tokens, the same as the preview. Tencent Cloud TokenHub sells it at half price until December 31, 2026, and WorkBuddy and CodeBuddy are free from November 4 to 15. A cheaper Hy4 Flash is "scheduled for release soon".

For another large open release this autumn, see how Mistral published the weights of Mistral Large 4, a model with 1 trillion parameters. If you build on Google's models instead, check the Gemini API changes due on December 1.