Tencent Releases Hy4, Adding Visual Understanding and Computer-Use Capabilities

Tencent released Hy4 on November 4, adding visual understanding and computer-use capabilities to the model. Compared with Hy4 Preview, the update’s technical focus is enabling the model to read images, analyze charts, and then use tool calls to perform tasks. Tencent says the model received targeted training on professional software such as Blender and CAD, and can interact with command-line and graphical interfaces. However, the announcement did not provide reproducible success rates for computer use. Tencent announcement

The developer interface has specific published details. TokenHub’s documentation lists hy4 as the production model identifier and supports three protocols: OpenAI Chat Completions, Responses, and Anthropic Messages. It also provides examples for tool calling, structured output, and image understanding. The context window is 1024k tokens, with a maximum input of 960k and a maximum output of 64k. Engineering teams therefore need to budget input and generation separately; the entire context window cannot be treated as available input capacity. TokenHub API guide

A third-party service has also confirmed availability. OpenRouter lists the release date as November 4 and the model identifier as tencent/hy4. Its currently displayed half-price rates are $0.417 per million input tokens and $1.251 per million output tokens. This gives existing compatible clients a direct way to test the model, but prices and provider routing should be checked against the settings of actual requests. Latency and throughput shown on the platform page should not be assumed to apply to long-context, multi-step tool tasks. OpenRouter model page

For agent developers, the key question to test is whether the model can consistently complete the cycle of understanding a screen, choosing a tool, and checking the result—not just how it performs on a single-turn answer. A useful evaluation would replay existing Chinese-language charts, cross-document tasks, and interface workflows, while tracking completion rates, error recovery, and cost per task. These are suggested tests based on the new capabilities, not confirmed performance gains.

Deployment boundaries still need attention. Although Tencent’s announcement positions Hy4 as an open-source model, it says the model weights and system card will be released in the coming days. What is confirmed so far is the availability of the hosted API; this does not establish that the production release already has a downloadable, reproducible self-hosting option. Follow-up checks should cover the weights’ license, inference framework support, and what the system card discloses about computer use and safety evaluations. Tencent announcement