Tencent’s November 4 announcement of Hy4 puts visual understanding and computer use alongside the text and coding tasks associated with large language models. The company describes a system intended to interpret visual information and work with both command-line tools and graphical interfaces.
Tencent lists creative and engineering environments including Blender, CAD software, Unreal Engine, Unity and Godot as examples of applications for the model. Those are the company’s descriptions of intended capabilities, not a demonstration that Hy4 can reliably complete any particular project. We have not tested it.
What is actually available?
Tencent describes Hy4 as available through its hosted services and API, while saying the final model weights and a system card are planned. When we checked Tencent’s public Hugging Face account on November 8, we found Hy4-preview and Hy4-preview-FP8. A preview is useful to investigate, but it should not be confused with verified availability of the final release.
The preview model card gives its own architecture and license information. Those figures and terms should not automatically be assigned to final Hy4. Tencent’s announcement also uses wording about Hy4 Flash that appears inconsistent about its release timing; we have not independently established its availability.
Why it matters
Multimodal computer use could make assistants useful in applications where the task is visible on screen, rather than fully described in text. But a model that can see and issue actions also needs clear permissions, user review and practical limits on what it can change. Benchmark scores alone cannot establish that such an agent is dependable or safe in a working project.
Our view: Hy4 is a development to watch, especially for creative tools. The next meaningful checks are the final system card, exact weight and license terms, and independent tests of real tasks. Until then, treat claims about performance and local hardware fit as unverified.
Independent reporting based on the linked sources. Product and performance claims are attributed to their publishers unless noted.
