Tencent’s Hy4 adds vision and computer use, and the weights are not out yet
The company’s 4 November release says Hy4 can read images and operate a command line or a graphical interface, including Blender and CAD, and that weights and a system card are days away. TokenHub pricing stays about $0.90 per million input tokens, with half off through 31 December.
Feature3 minIn-depth
James Okonkwo3 min read
- Tencent
- Hy4
- Open weights
- Agents

Photo: TencentSource page
Tencent released Hy4 on Wednesday and said the model, for the first time in that line, can read images and operate a computer. The newsroom post is dated 4 November. Visual understanding covers fine-grained images, charts, and visual mathematics, in the company’s description. Computer use covers a command line and a graphical interface, with extra training on Blender and CAD so the model can plan work that runs for a long stretch and across more than one program. Tencent says Hy4 now ranks among the leading open-source models. It also says the weights and the system cards will come in the coming days, with the release notes aimed at cybersecurity and biological safety. Until those files are public, the ranking is the company’s description of a model outsiders cannot yet download.
The model is already in Tencent products: Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom’s assistant Dayuan, and ima. The post says the API is also going out through OpenRouter, Hermes, and OpenCode, and that QQ, Honor of Kings, Game of Peace, WeRead, and Tencent News are still being connected. The price, Tencent says, is unchanged from Hy4 Preview despite the gains: about $0.90 per million input tokens, $2.70 per million output tokens, and $0.045 per million cached input tokens. From 4 November through 15 November, WorkBuddy and CodeBuddy are free. On TokenHub, Hy4 is half off through 31 December 2026. A cheaper Hy4 Flash is described as coming soon, without a date. Tencent Cloud’s API guide, timestamped 12:10:59 on 4 November, lists the model id as hy4, with a context window of 1 million tokens, a maximum input of 960,000, and a maximum output of 64,000. The same page says Hy4 accepts images on the Chat Completions, Responses, and Anthropic-style APIs.
Scores from Tencent’s own tests
The comparisons in the post are Tencent’s. With its financial-technology group, the Hy team ran a blind test of 579 tasks and 213 financial professionals, 153 of them from outside institutions and about 70 percent of them with more than five years in the field. Hy4 scored 3.04. DeepSeek-V4.1-Flash-1017 scored 2.97. Kimi K3.1 Pro and GLM-5.3 each scored 2.68. On an internal investment-research benchmark run by the financial AI company Rabyte, Tencent says Hy4 gained 12.03 percentage points on Hy4 Preview. In internal checks, the hallucination rate fell from 12.95 percent on the preview to 7.2 percent, and errors from unsupported assumptions fell from 23.1 percent to 5.6 percent.
On WorkBuddy, Tencent reports an average score of 83.18 and a pass rate of 96.15 percent, and first place in nine of twelve scenarios in the mainland Chinese market. Against Hy4 Preview, it reports a 19.1 percent gain on a GameDev benchmark and 9.8 percent on GameLogic, after work with Tencent’s games group on Unreal Engine, Unity, and Godot, including multiplayer. The post also names a harness, Hyra 2.0. Using it, Tencent says, Hy4 optimized pre-training infrastructure for smaller models of the same architecture, and the team has open-sourced two models built with Hy4 and that harness: HyCell, for single-cell representations, and HyPDE, for physics fields. Those two are the weights the post says are already public. Hy4’s own weights are the ones still described as days away.
On safety, the post says Hy4 was trained on defensive security data, with proof-of-concept work and vulnerability repair in scope and offensive exploitation material left out of the training set. Tencent’s Xuanwu Lab and Keen Security Lab reviewed cryptographic reverse engineering, binary deobfuscation, and vulnerability exploitation, and the post says they found no significant concerns about offensive cybersecurity capability in the areas they tested. On biology, Tencent says internal reviewers found strong foundational knowledge and more limited skill at open-ended experimental design than at frontier models. Both sentences are the company’s summary of its own review.


