YFarmX logoYFarmX

Tencent

Hy4

Tencent’s model for reasoning and computer tasks

Released 4 November 20261 min readLarge Language ModelsLast updated:

Conceptual collage of a detailed monitor, spreadsheet and circuit board with Tencent branding.

Key facts

4 Nov 2026Tencent release
Announced
Hy4production release
Model family
Multimodalincluding visual understanding
Inputs
7 Dec 2026CloudBase, Beijing time
Preview migration

Hy4 is Tencent’s production AI model for reasoning, coding, visual understanding and computer-use tasks. Tencent announced it on 4 November 2026.

Hy4 is Tencent’s AI model for reasoning, coding, visual understanding and tasks that involve using a computer. Tencent announced the production model on 4 November 2026. It follows the earlier Hy4 preview.

The production model has its own release identity

Hy4’s launch covers a later production model. The preview’s weights, licence and benchmark observations belong to that earlier release. Tencent’s production announcement says weights and system cards are planned; a release announcement by itself does not establish that those files are available to download.

Tencent describes visual understanding and computer-use capabilities. In an application, those capabilities depend on the tools and permissions provided to the model. Human review and access controls remain part of a safe deployment.

CloudBase is migrating the preview endpoint

CloudBase’s retirement notice sets 7 December 2026 at 00:00 Beijing time as the cutoff for its Hy4 preview endpoint. Unmigrated requests will automatically route to Hy4 and use Hy4 billing. Check request behaviour and costs before that service change.

The November service changelog records hosted model availability. Service access and downloadable weights have separate terms.

Hy4 flash is a smaller active model

Tencent announced Hy4 flash on 16 November for lower-cost, faster everyday workflows. The Hy4 flash guide covers its context window, architecture and time-limited offers. Compare the specific model on your intended workload, rather than treating a family name as a single performance score.