Tencent releases Hy4: 770B total parameters, 49B active, with native vision and computer use

Tencent has released Hy4 (混元Hy4), a model with 770B total parameters and 49B active parameters. It adds native vision and computer-use (CUA) support for the first time, according to Chinese-language outlet QbitAI (量子位), which ran hands-on tests on 2026-11-09. Tencent positions Hy4 among the strongest open-source models. Its weights are due to be open-sourced "in the near future."

What is confirmed

• Size: 770B total parameters, 49B active.
• New capabilities: native visual understanding and computer operation, both first-time features. Instruction following, long-document understanding and long-horizon agent ability were also upgraded.
• Training: the release follows a preview version. Tencent says it expanded pre-training and post-training data and scale, and co-designed the model with front-line business teams.
• Availability: Tencent Cloud TokenHub, WorkBuddy and CodeBuddy, Yuanbao, the Enterprise WeChat assistant, and ima. The API is also on OpenRouter, Hermes and OpenCode. QQ, Honor of Kings, Peacekeeper Elite, WeChat Reading and Tencent News are being connected.
• Free trial: available in WorkBuddy and CodeBuddy until 23:59 on November 15.
• Not yet disclosed: context window and per-token pricing.

Tencent's reported results

The figures below come from Tencent's internal testing unless noted.

• Hallucination: the rate fell from 12.95% in the preview to 7.2%. The speculative-error rate fell from 23.1% to 5.6%.
• Vision: Tencent says Hy4 can locate roof components of about 2 meters in an aerial image of 28亿 pixels (the source's figure, in Chinese numerals).
• Finance: training covered fund research, securities analysis, quant strategy, asset allocation, corporate finance, and tax and accounting. Tencent says it built thousands of workspaces and over ten thousand tasks. 213 experts took part in crowd testing, and seven in ten were from external top financial institutions with at least five years of experience. On 579 real financial tasks, Tencent says Hy4 ranked first among open-source models.
• Research: on the internal TenureTrail evaluation, covering 7 disciplines and 123 executable research tasks, Tencent says Hy4 showed research ability in mathematical construction and physics modeling.
• Biology: given 84 unordered 2D mouse-embryo slices containing 699万 single cells, with mirror-flip noise, offline and without registration tools or cell-type labels, Hy4 reportedly completed spatial alignment and 3D reconstruction in 1.5 hours with no human intervention. Tencent says the alignment precision exceeded the default workflow of a professional registration tool.
• Mathematics: working with a Peking University math team, Hy4 reportedly found a counterexample to the Méndez–Pinto–Villarreal conjecture, proposed a new condition to refine it, and had the related proof checked with Lean formal verification.

Hands-on tests

QbitAI tested Hy4 inside WorkBuddy. Given a design screenshot, it built a retro music player. It measured layout details, synthesized three test tracks with Web Audio, and ran a headless Chrome check in a sandbox. The first pass cleared 51 test items with 48 passing. After it fixed an edge case at the end of the playlist, all 59 tests passed with zero console errors.

It also built a navigable 3D amusement park from a floor plan and produced a 16-page product deck from supplied materials. When text overflowed on page 4, it restructured the layout on its own. For a re-template to a dark theme, it tested one page before applying the change to the rest. It built a batch image-processing tool and used its CUA ability to draw in JS Paint in Chrome by mouse.

Training also targeted front-end work (complex interaction, state management, front-end and back-end integration) and professional software such as Blender and CAD. Tencent says it optimized the search pipeline for time-sensitive retrieval, multi-source integration and factual consistency.

Tencent says Hy4 flash, aimed at very low inference cost, on-device use and high-concurrency agent orchestration, is on the way.

What this means

Hy4 is the first Hy model with native vision and CUA, and QbitAI's tests put those capabilities to work in a loop of building, testing and fixing. The hallucination figures are Tencent's internal numbers against its own preview, not independent benchmarks. The open-source claim cannot be checked until the weights ship, and the context window and pricing are still undisclosed. Developers can try it now through the free trial window or the OpenRouter, Hermes and OpenCode APIs.