Tencent's Hy4: how to test the open-source model with vision

In this article you will see:

• What Hy4 is and what changed compared to Hy4 preview
• How to test Hy4 today, locally or via API
• Where the model excels and where it still loses to closed models
• Frequently asked questions about Tencent's Hy4

What is Tencent's Hy4

Hy4 is the stable version of the Hy family from Tencent Hunyuan. Compared to Hy4 preview, launched on August 28, 2026, the new model expanded pre-training and post-training, introduced vision for the first time and improved instruction following, long context understanding and long horizon tasks. The model is open-source and is available on Hugging Face.

The architecture is a Mixture of Experts, MoE, with 770 billion parameters in total and 49 billion activated per token. There are 78 layers, 77 of which use MoE with 256 routed experts and 1 shared expert, activating the top 8 per token. The context window is 1 million tokens, at the same level as frontier open-weights models from Europe and China.

How to test Hy4 today

Hy4 is available in three ways.

First, via hosted API. Tencent makes the model available on TokenHub, OpenRouter, Hermes and OpenCode. The endpoint is compatible with the OpenAI API, so any client that already calls Chat Completions works by swapping the base URL and the model.

Second, self-hosted. The weights are on the repository Tencent-Hunyuan/Hy4-preview on Hugging Face. The official documentation recommends serving with vLLM or SGLang and calling the local endpoint with the same OpenAI interface. The recommended parameters are temperature 0.9 and top_p 1.0, with reasoning mode set to "high" by default for complex tasks.

Third, inside Tencent products. The model is already in WorkBuddy, CodeBuddy, Yuanbao, ima and the intelligent assistant of WeCom. Tencent offered limited free use in WorkBuddy and CodeBuddy between November 4 and 15, 2026.

For programmers, the fastest path is the API on OpenRouter or the Tencent endpoint, testing a code prompt with long context and an attached image to validate the vision part. If you already run local models, it is worth comparing Hy4 with the options in the guide for open source AI assistants to run locally and with the flow of Claude Code in VS Code.

What Hy4 does well

The model was trained with data from Tencent engineers, game developers, financial analysts and security specialists. The result shows up in long software tasks: planning, debugging and verification of changes that extend over several steps. In internal tests, Hy4 improved 19.1% in GameDev and 9.8% in GameLogic compared to the preview.

Vision is new. The model accepts images, charts and documents, and does diagram analysis and visual mathematical reasoning. Computer use is also new: Hy4 can operate graphical and command line interfaces for day to day tasks.

In a blind test with 163 internal specialists on 203 engineering tasks, Hy4 was slightly ahead of GLM 5.3, 2.99 versus 2.92 average, and Kimi K3, 2.99 versus 2.94. In finance, 213 specialists evaluated 579 real tasks and Hy4 ranked first among open-source models.

Where Hy4 still falls short

Tencent is explicit about the limitations. The model still spends more time than necessary reasoning about complex tasks and tends to over-check its own work. In frontier math, long-horizon development and difficult visual questions, the gap to the best closed models is still large.

The company also says the model has real headroom in both pre-training and post-training, and plans to launch Hy4 flash soon with better cost efficiency and speed. The expectation is rapid iteration, as happened with Hy4 preview.