Tencent releases Hy4 with vision and computer use, weights promised in coming days

A better model, a thinner release

Tencent offers Hy4 now through Tencent Cloud and OpenRouter at $0.834 per million input tokens and $2.501 per million output tokens.

For teams that picked the August preview because they could download it, the successor does not offer that yet. Tencent says it will publish the open weights and a system card in the coming days, but it has not named a licence. OrcaRouter noted on 4 November that the weights are announced as a plan and the model card does not exist.

Tencent did not publish a parameter count in its announcement. TechNode reports that Hy4 has 770 billion parameters.

What is new is capability. Hy4 reads fine details in images, analyses complex charts and solves visual maths problems. Its computer-use mode drives command-line tools and graphical interfaces. Tencent says it trained the model on professional software including Blender and CAD. That training lets it plan and finish longer multi-step tasks. That part is testable through the API today.

Tencent said it limited cybersecurity training data to defensive material, filtering out content on exploit development, reverse engineering and penetration testing. It said it recognizes safety risks from advanced cyber and biological capabilities.

The release lands in a crowded field. The South China Morning Post reported on 4 November that Tencent is pitching Hy4 against open-weight rivals including DeepSeek, Z.ai and Alibaba.

What the preview gave away

Hy4 Preview shipped on 28 August 2026 under an Apache 2.0 licence. Its 770 billion parameters were downloadable from Hugging Face, and it carried a benchmark appendix of eight rows of published numbers. Tencent has not published a licence for the new build, and it released no benchmark appendix.

The preview's published figures were 770 billion total parameters, 49 billion activated per token, 78 layers and a 1,048,576-token context. None of that paperwork accompanies the new model. Tencent has not published an architecture description for it. It frames the change as scale in pre-training and post-training rather than a new design. Whether the vision encoder is a separate model stitched on is an inference, not a disclosure.

The preview's model card was unusually candid about two problems: it reasoned longer than necessary on hard tasks and over-verified its own answers. Tencent says the new build shortens the reasoning and improves instruction following. It published no before-and-after token counts.

OrcaRouter's 4 November comparison put it plainly: in the Hunyuan line the version number went up and the openness went down.

The numbers Tencent did publish

Tencent reports gains of 19.1% on its GameDev benchmark and 9.8% on GameLogic compared with Hy4 Preview. It says the hallucination rate fell to 7.2% from 12.95% after it reworked the model's search features. OrcaRouter's comparison gives different figures for the same metric, 15.76% to 5.45%, and says the methodology is unpublished. Errors caused by unsupported assumptions dropped from 23.1% to 5.6%, Tencent says. Those are Tencent's own tests, measured against its own previous model.

In a blind evaluation with 213 finance professionals, 153 of them from external institutions, across 579 real tasks, Hy4 scored 3.04, ahead of DeepSeek-V4.1-Flash at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68 each. On WorkBuddy, Tencent's AI work assistant, the model posted an average score of 83.18 and a 96.15% pass rate. Tencent says it came first in nine of the 12 scenarios it tested in mainland China. It also reports a 12.03 percentage point gain over the preview on iRaB, an investment research agent benchmark run by financial AI company Rabyte.

ShiftDelete reported that none of these figures had been checked independently.

Price held, discounts stacked

Tencent Cloud TokenHub is offering Hy4 at 50% off through 31 December 2026, and Token Plan subscriptions carry a 75% discount on credit consumption through 18 November.

A cheaper Hy4 Flash version will launch soon, Tencent says. It will inherit vision understanding and agent functions.

What remains open

Tencent has acknowledged a gap with top proprietary models. It names research-level math, long-horizon coding and advanced visual question answering as the weak spots.

Until the weights appear, Hy4 is available only through APIs and Tencent's own apps.