Tencent launches full Hy4 with vision and computer use; weights to follow

What this is about

On November 4, 2026, Tencent unveiled the full version of its Hy4 language model. The preview came out as an open-weight model on August 28, 2026. The new features are image understanding and computer use, meaning the ability to operate software through the command line and graphical interfaces. For now, the model runs only as a hosted service. Tencent says the weights and a comprehensive system card (a document covering safety evaluations and technical details) will arrive "in the coming days". According to techbooky, it is not yet confirmed that they are available.

Technology and what's new

Hy4 is a Mixture-of-Experts model. Only part of the network is activated for each token, so compute costs track the 49 billion active parameters, not the 770 billion total. The context window is about 1 million tokens. Tencent significantly expanded pre-training and post-training. It cites gains over the preview in instruction following, long context, long agentic tasks and deep research. In the preview, Tencent itself had criticized excessive overthinking and over-checking.

With image understanding, Hy4 is meant to handle fine image details, complex charts and visual math problems. For computer use, Tencent trained specifically on Blender, CAD and workflows spanning multiple applications. According to SiliconANGLE, Tencent also used its Hyra 2.0 agent system so the model could optimize the pre-training infrastructure of smaller models with the same architecture (recursive self-improvement).

Numbers from Tencent's own tests

All of the following figures come from the vendor and have not been independently verified.

SiliconANGLE points out that the baseline values for the game benchmarks remain internal. The game work covered Unreal Engine, Unity and Godot, and developers from the Tencent Interactive Entertainment Group helped with training data. On the internal TenureTrail test (123 research tasks across seven subjects), Tencent says Hy4 led among open-weight models. A gap to the best proprietary models remains in research-level math, long-horizon coding, self-improvement and demanding visual questions.

Availability and pricing

Hy4 is available through Tencent Cloud TokenHub, OpenRouter, CodeBuddy, WorkBuddy, Yuanbao, ima and the WeCom AI assistant. According to SiliconANGLE, Hermes and OpenCode are included as well. Unite.AI lists $0.834 per million input tokens, $2.501 for output and $0.042 for cached input. Other sources round these to about $0.90, $2.70 and $0.045. That matches the preview pricing. OpenRouter lists 1,048,576 tokens of context and 65,536 tokens of output, with text and image input and text output. Tencent Cloud is the only provider there. According to techbooky, TokenHub is 50% off until December 31, and WorkBuddy and CodeBuddy are free from November 4 to 15.

Safety

For cybersecurity data, Tencent limited training to defensive proofs of concept and patches. It filtered out exploit development, reverse engineering and penetration testing. Tencent's Xuanwu and Keen Security labs ran red-team audits and, according to Tencent, saw no elevated risk. On the Hy-BioExperimentBench biology test, Hy4 scored 30.4 out of 100. Internal reviewers saw no meaningful lowering of the barrier to biological misuse.

What you need to try it

Nothing runs locally yet because the weights are missing. Once they arrive, the 770 billion parameters mean you'll likely need server hardware with a lot of memory. Tencent plans to provide support for vLLM and SGLang. Until then, you can test Hy4 through the API on OpenRouter or in Tencent's products.

Our take

Hy4 shows how Chinese vendors are pushing open models toward agents and office work. The price stays low. On the critical side, the full version can't be downloaded yet, the license terms are unknown (tech-insider), and almost all benchmarks are company-run. Tech-insider notes there is no verifiable benchmark table. Computer use also brings risks around permissions and data access, which users have to secure themselves. Independent tests will have to show whether Hy4 really belongs among the leading open models.