Tencent launches Hy4 with visual reasoning and computer use, with weights still planned

Tencent announced a new Hy4 model milestone.

Tencent announced the final Hy4 release on November 4. It adds visual understanding and computer-use capabilities and is intended for tasks involving charts, software interfaces and extended sequences of work. Tencent said it planned to publish model weights and system cards in the coming days. The launch is a product milestone. The separate checkpoint release would establish which weights users can download and the terms under which they can run them. [Tencent-issued announcement](https://en.prnasia.com/releases/apac/tencent-launches-hy4-joining-the-ranks-of-the-world-s-leading-open-source-ai-models-554689.shtml)

How the new capabilities show up in products

Tencent describes a model that can use command-line and graphical interfaces as well as interpret images. In WorkBuddy, Tencent says it can work with information from connected WeCom workflows such as messages, documents and meetings. Tencent also describes specialist training on applications such as Blender and CAD tools. Those examples indicate where the company expects the model to help. They do not establish that every application or action is available to every customer. [Tencent's product examples](https://en.prnasia.com/releases/apac/tencent-launches-hy4-joining-the-ranks-of-the-world-s-leading-open-source-ai-models-554689.shtml)

A workflow that summarizes documents can be evaluated with a fixed set of source records. A workflow that changes a file needs an additional check on what the system is allowed to modify and how to undo an incorrect action. Visual understanding can help identify the relevant control, but the surrounding application still determines whether the model can use it and whether a person must approve the action.

What to distinguish when evaluating the launch

Tencent reports results from a financial evaluation involving 213 professionals and 579 tasks. Its release also cites comparisons with Hy4 Preview in specialized benchmarks. These figures are useful for seeing what the company chose to measure. A team evaluating a specific workflow would need to inspect how the tasks were selected, what qualified as a pass and whether the model is working under comparable access constraints. [Tencent's evaluation description](https://en.prnasia.com/releases/apac/tencent-launches-hy4-joining-the-ranks-of-the-world-s-leading-open-source-ai-models-554689.shtml)

The announcement establishes the new release and the intended weights publication. Those are two separate events, with different evidence for product users and teams planning to self-host.