Tencent launches Hy4 with vision and computer use, but the open weights are still to come

Tencent on Wednesday announced Hy4, the full release of its flagship large language model, two months after it published open weights for a preview version. The company says Hy4 adds two capabilities the Hy line did not have before: image understanding and computer use. The announcement came through a Tencent press release distributed by PR Newswire at 19:13 China time (08:13 in Buenos Aires).

The headline framing is that Hy4 "ranks among the world's leading open-source models". There is a catch worth stating up front: the Hy4 weights are not public yet. Tencent says it plans to release model weights and system cards "in the coming days". Until then, Hy4 is a hosted model that you reach through Tencent's apps and APIs, and every performance figure in the announcement comes from Tencent or its partners.

What Tencent says is new

• Vision. Tencent describes visual understanding as a first for the model, aimed at fine-grained image processing, complex chart analysis and mathematical visual reasoning.
• Computer use. Also a first, according to the company: Hy4 is meant to operate both command-line and graphical interfaces. Tencent says it received specialized training on professional software such as Blender and CAD tools for long-horizon tasks.
• Fewer hallucinations, by Tencent's own measurement. In internal evaluations, the hallucination rate fell from 12.95% in Hy4 Preview to 7.2%, and errors from unsupported assumptions fell from 23.1% to 5.6%. Tencent did not publish the evaluation set.
• Coding and game development. Tencent reports gains of 19.1% on a GameDev benchmark and 9.8% on a GameLogic benchmark versus the preview, plus integration work with Unreal Engine, Unity and Godot.

The benchmark claims, read carefully

The most specific comparison in the release is a blind evaluation for financial work, run by the Tencent Hy team with Tencent Financial Technology. Tencent says 213 financial professionals, 153 of them from external institutions, rated outputs on 579 real-world tasks. The reported scores: Hy4 3.04, DeepSeek-V4.1-Flash-1017 2.97, Kimi K3.1 Pro 2.68 and GLM-5.3 2.68.

That is a vendor-run study comparing Hy4 only against other open models, on a domain Tencent itself chose, and the scale of the scores is not explained. It is useful as a signal of where Tencent is aiming, not as independent evidence. The same applies to an office-productivity figure the company gives for its WorkBuddy agent (an average score of 83.18 and a 96.15% pass rate) and to a 12.03-point gain on an investment-research benchmark run by the financial AI company Rabyte.

Price and availability

Tencent says Hy4 keeps the preview's pricing: about US$0.90 per million input tokens, US$2.70 per million output tokens and US$0.045 per million cached input tokens. It is also running two promotions: Hy4 is 50% off on Tencent Cloud TokenHub through December 31, 2026, and the WorkBuddy and CodeBuddy tools are free from November 4 to 15.

The model is live in Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, the WeCom assistant and ima. Tencent says API access is "being made available" through third-party platforms including OpenRouter, Hermes and OpenCode, so availability there may lag the announcement. A cheaper Hy4 Flash variant is described as "scheduled for release soon", although the section heading in the same release says it was "recently released". We could not confirm that Flash is available yet.

What we know about the architecture, and what we don't

The release does not give Hy4's size or context window. The best available reference is the Hy4 Preview model card on Hugging Face, published on August 27-28 under the Apache 2.0 license: a Mixture-of-Experts model with 770B total parameters and 49B active per token, a 1M-token context length, and sparse attention borrowed from the DeepSeek line. Tencent has not said whether the final Hy4 keeps that architecture, nor what license the final weights will carry.

The preview card also listed known issues: the model spent longer than necessary reasoning through complex tasks and tended to over-verify its own work. Today's release does not say whether those were fixed, beyond the hallucination numbers above.

Safety claims

Tencent puts unusual emphasis on cybersecurity and biosafety in this release. It says Hy4 was trained on defensive security data, with exploitation-related offensive material excluded, and that its Xuanwu Lab and Keen Security Lab found "no significant concerns" about offensive cyber capabilities in the areas they evaluated. On biology, Tencent says the model has strong foundational knowledge but is more limited than frontier models at open-ended experimental design. These are internal assessments; the promised system cards will be the first chance to see the methodology.

Tencent also says it open-sourced two scientific models built with Hy4 and its Hyra 2.0 harness, HyCell for single-cell representations and HyPDE for physics fields.

Our read

For developers, the interesting part of Hy4 is not the leaderboard language but the combination: a model in the 770B-class MoE family, priced under a dollar per million input tokens, that now accepts images and is trained for agentic computer use, with a stated plan to publish weights. If the weights ship under a permissive license like the preview's, Hy4 becomes a serious candidate for teams that want a self-hostable or cheap hosted model for multimodal agent work. Until they ship and outside evaluators test it, treat the performance claims as Tencent's.