Tencent Hy4 launches: 770B model adds vision and Computer Use, with open weights to follow

Tencent Hunyuan released the official version of Hy4 on November 4, 2026, turning the preview it shipped in August into its finished flagship. The headline change is not scale alone: for the first time in this model line, Hy4 can see. It adds visual understanding and Computer Use to a 770-billion-parameter mixture-of-experts model built, in Tencent's framing, for real productivity work rather than chat.

That matters because the preview was text-only. Compared with Hy4 preview, Tencent says Hy4 significantly expands pre-training and post-training scale and improves instruction following, long-document understanding, long-horizon agent work and deep research. The model weights are not out yet — Tencent says they will be open-sourced soon, alongside a model system card — so for now the evidence is Tencent's own reporting, not independent testing. Read the release that way: a strong vendor claim, with the weights still to arrive.

Vision, then hands on the keyboard

Hy4's visual understanding combines perception, reasoning and tool use, according to reporting on Tencent's announcement. The examples given are granular image processing, complex chart analysis and mathematical reasoning over images. Computer Use goes a step further: the model can operate a computer by combining command-line and graphical interfaces, with specialist training for professional software such as Blender and CAD and for long tasks that cross applications. That combination — look at a chart, reason about it, then act on it in software — is the practical core of this release.

Tencent has also tuned Hy4 for front-end and game development, working with game developers from Tencent IEG on game features, gameplay logic, state understanding and asset handling, with adaptation to the UE, Unity and Godot engines. Against the preview, Tencent reports GameDev evaluation scores up 19.1% and GameLogic up 9.8%. The same visual capability feeds office work: Tencent says Hy4 can collaborate with users in real time on Word, Excel and PowerPoint documents.

The finance test Tencent is leading with

The most concrete numbers in the release come from finance. Tencent built thousands of workspaces and tens of thousands of training tasks around fund research, securities analysis, quantitative strategy and related work, then ran crowd-testing with Tencent Financial Technology (FiT): 213 experts, 153 of them from leading external financial institutions and 70% with more than five years' experience, over 579 real tasks. Hy4 scored 3.04 in that test, ahead of Kimi K3.1 Pro at 2.68, GLM-5.3 at 2.68 and DeepSeek-V4.1-Flash-1017 at 2.97. Those are Tencent's figures from its own test design, and the comparison models' scores come from the same exercise — useful as a signal, not as an independent benchmark.

In office scenarios, Tencent reports a WorkBuddy evaluation average of 83.18 with a 96.15% task pass rate, ranking first in 9 of 12 scenarios domestically. As with the finance test, that is a vendor-run evaluation in Tencent's own products, where Hy4 can pull together chat records, email, documents, schedules and meetings through the WeChat Work suite. The honest version: Hy4 looks tuned for the exact workflows Tencent sells, and the numbers say most about those workflows.

Research, and a model helping build models

For scientific research, Tencent points to TenureTrail, an internal evaluation covering 7 disciplines and 123 executable research tasks, where it says Hy4 ranked first among open-source models by combining complex reasoning, scientific programming and long-horizon execution. Tencent also says Hy4 is already being used in model development itself: combined with the Hyra 2.0 scaffolding, it helped optimise pre-training infrastructure for smaller models and automatically train a model that outperforms the official version in multiple evaluations. Two resulting foundation models, HyCell for single-cell characterisation and HyPDE for physical fields, have been open-sourced as reusable tools.

Safety claims, stated plainly

Tencent is publishing the caveats alongside the claims. It says Hy4 is one of the strongest open-source models, but still has a significant gap to the top proprietary models, especially on hard internal test sets in advanced maths, long-horizon development, model iteration and visual question answering — and that it will keep expanding training scale to close it. On security, Tencent says Hy4's training included only defensive cybersecurity data, such as proof-of-concept generation and vulnerability remediation, excluding offensive content like exploitation and penetration testing, with audits by security experts from Tencent's Xuanwu and Cohen labs finding no significant offensive-capability risk. On biosafety, it reports a Hy-BioExperimentBench score of 30.4 out of 100 for open experimental design, describing a significant gap to cutting-edge models, and says no significant lowering of the threshold for biochemical-hazard behaviour was found. A fuller safety picture is promised with the system card when the weights are released.

Where to try it, and what it costs

Hy4 is already live in Tencent Cloud TokenHub, WorkBuddy and CodeBuddy, Yuanbao, ima and the WeCom intelligent assistant Dayuan, with APIs integrated into OpenRouter, Hermes and OpenCode, and further Tencent products such as QQ, Honor of Kings, WeChat Reading and Tencent News integrating it. Pricing is unchanged from the preview: 6 yuan per million input tokens, 18 yuan per million output tokens and 0.3 yuan per million tokens for cache hits. WorkBuddy and CodeBuddy are offering some free use from November 4 to November 15, and Hy4 is charged at a 50% discount on TokenHub until December 31, 2026, based on the published price. Hy4 Flash, a cheaper and faster sibling carrying over Hy4's visual and agent capabilities, is to be released soon.

For background, Tencent previewed the general release of Hy4 and Hy4 Flash on October 29 at its Global Digital Ecosystem Summit, and this site covered that preview alongside WorkBuddy's global launch. The November 4 release is the delivery of that preview: the same model line, now finished, visual, and able to use a computer. Whether it earns the "strongest open-source" label will be decided once the weights are out and independent evaluations — not Tencent's — have had their turn. The comparison to watch is not the preview, but DeepSeek's V4.1-Flash and the other open-weights models, such as Xiaomi's MiMo-V2.6, that Hy4 was measured against in Tencent's own tests.