Tencent Announces Hy4, Claims Top Tier Open-Source LLM Status With Weights to Follow

What Tencent announced

Tencent has unveiled Hy4, the newest version of its large language model, and says the system now performs at the level of the best openly available AI models. According to the company's announcement, the model weights and accompanying system cards will be published in the next few days, so at announcement time the open-source release itself was still pending rather than complete.

The release adds visual understanding to the model line for the first time. Tencent says Hy4 can handle detailed image work, complex chart interpretation and reasoning over visual mathematics. It also gains computer-use features, meaning it can drive both command-line and graphical interfaces to carry out work, and it received targeted training on professional tools such as Blender and CAD to support longer, multi-step jobs. The company also reports broad gains in instruction following, long-context comprehension and long-horizon agentic tasks.

The numbers, and who produced them

Nearly all the performance evidence comes from Tencent's own announcement or evaluations the company ran itself, so these figures are vendor claims until third-party reviews appear.

The headline result is a blind comparison organised by the Tencent Hy team together with Tencent Financial Technology. It put 579 real financial tasks in front of 213 professionals, 153 of them from outside institutions and roughly 70 percent with more than five years in the industry. Hy4 scored highest overall at 3.04, ahead of DeepSeek-V4.1-Flash-1017 at 2.97, while Kimi K3.1 Pro and GLM-5.3 both landed at 2.68. Separately, financial AI firm Rabyte measured a 12.03 percentage-point jump over the earlier Hy4 Preview on its Investment Research Agent Benchmark.

Other self-reported figures cover office and developer scenarios. On Tencent's own WorkBuddy suite, the model averaged 83.18 with a 96.15 percent pass rate and placed first in nine of twelve scenarios within the mainland Chinese market. Game development improved 19.1 percent on a GameDev benchmark and 9.8 percent on GameLogic compared with the preview, aided by work with Tencent's games division on Unreal Engine, Unity and Godot. On information retrieval, the claimed hallucination rate fell from 12.95 percent to 7.2 percent, and errors from unsupported assumptions dropped from 23.1 percent to 5.6 percent.

Availability

Hy4 is already live across Tencent products including Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom's AI assistant Dayuan and ima, and its API is being offered through OpenRouter, Hermes and OpenCode. Tencent says further integrations are planned for QQ, Honor of Kings, Game of Peace, WeRead and Tencent News, and that a lower-cost Hy4 Flash variant has also shipped.

Safety and research commitments

The announcement emphasises safety work ahead of the weight release. For cybersecurity, the model was trained on defensive material such as proof-of-concept development and vulnerability fixing, and Tencent says offensive exploitation content was deliberately excluded from the training data. Its Xuanwu Lab and Keen Security Lab reviewed areas including cryptographic reverse engineering and binary deobfuscation, and reported no significant concerns about offensive capability in the areas they assessed. On biosafety, internal experts judged the model strong on foundational biology but weaker than frontier systems at open-ended experimental design.

The team is also pursuing recursive self-improvement using a harness it calls Hyra 2.0, and has released two scientific foundation models built with it: HyCell for single-cell representations and HyPDE for physics fields.

Why it matters

Hy4 lands in an increasingly crowded field of Chinese open-weight models, with DeepSeek, Kimi and GLM named as the direct competitors in Tencent's own evaluation, and the company is claiming the top spot among them. If independent testing backs that up, a downloadable model at this level would pressure both Western closed labs and rival open-source projects. The announcement also shows what vendors now treat as table stakes for shipping weights: vision, computer use, agentic execution and a public safety review. The caveat is that nothing is verifiable yet. The weights have not been released, and every number above originates from Tencent itself.