Tencent launches Hy4 model with 770B parameters
Hy4 supports a one-million-token context and adds image understanding and computer control; full weights and a system card are still to come.
Tencent announced the launch of Hy4, its next-generation AI model, on November 4, according to aitimes.com. The release follows Hy4 Preview, which the company put out in August. Tencent has not yet published the full model weights; it says they will be distributed soon alongside a system card containing safety evaluation results.
Hy4 uses a mixture-of-experts architecture with 770 billion total parameters and about 49 billion activated during inference. It supports a context of up to one million tokens and introduces image understanding for the first time, including complex chart analysis and mathematical visual reasoning. The model also adds computer operation abilities covering command-line and graphical user interfaces, and was trained on professional software such as Blender and CAD, with improved frontend development and game creation performance.
Benchmark results
In an evaluation of 579 practical tasks by 213 financial experts, 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. In WorkBuddy's 12 evaluation items, it ranked first among Chinese models in 9, with a 96.15% overall task pass rate. Tencent's own evaluation put Hy4's hallucination rate at 7.2%, down from 12.95% in the preview.
The picture was more mixed in external testing. In Vals AI's evaluation published on November 4, Hy4 ranked 13th of 33 models with 59.35% overall accuracy and 4th among open-weight models, trailing Kimi K3.1 Pro (62.52%), DeepSeek V4.1 Flash (61.26%) and GLM-5.3 (59.73%). Hy4 scored 39.39% on Terminal Bench 4.0, ranking 2nd among open-weight models.
Tencent said its training data excluded attack material usable for breaking into systems or exploiting vulnerabilities, focusing instead on defensive data such as vulnerability analysis and patching. Its Xuanwu Lab and Keen Security Lab found no notable risk related to attack capability in evaluations of cryptanalysis, binary deobfuscation and vulnerability exploitation.
Hy4 is available in Tencent Cloud, CodeBuddy, WorkBuddy and Yuanbao, with API pricing kept at the same level as the preview. WorkBuddy and CodeBuddy offer a set amount of free usage until the 15th. Tencent also plans to launch a lighter model called Hy4 Flash, with improved inference speed and cost efficiency.
Quick answers
Where can I use Tencent's Hy4?
Hy4 is available in Tencent Cloud, CodeBuddy, WorkBuddy and Yuanbao, with API pricing at the same level as the preview.
Are Hy4's model weights available?
Not yet. Tencent says the full weights will be distributed soon together with a system card containing safety evaluation results.
How does Hy4 score on Vals AI's evaluation?
It ranked 13th of 33 models with 59.35% overall accuracy and 4th among open-weight models, in results published on November 4.