Tencent has released the Hy4 preview model today, with a total of 770B parameters and 49B activated parameters, and a context length of up to 1M. The official stated that the model has significantly expanded in size, context, and data scale, and the simultaneous advancement of pre-training and post-training has led to another leap in intelligence level, firmly securing its position among the top open-source models. The model is now open-sourced and has been synchronized on HuggingFace, GitHub, ModelScope, and Gitcode, and has also been integrated into Tencent Cloud TokenHub and OpenRouter.

image.png

Productivity scenarios have been fully upgraded, outperforming GLM-5.3 and Kimi K3 in blind tests

Based on high-quality data co-built with experts in software engineering, gaming, finance, and security within Tencent, Hy4 preview has made significant progress in real tasks: it strengthens long-term development understanding, planning, and verification on the software engineering side; it can complete full delivery from data processing to document table presentation in office analysis; it can generate playable prototypes with one requirement in game development; and it has made progress in research scenarios such as molecular dynamics and condensed matter physics.

image.png

Tencent organized 163 internal experts and 203 engineering tasks for blind testing, and Hy4 preview achieved an average score of 2.99 (out of 4.00), slightly surpassing GLM-5.3's 2.92 and Kimi K3's 2.94. Along with Hyra, it has also made breakthroughs in the century-old geometric problem of three-dimensional Blaschke–Lebesgue, advancing the volume lower bound from 0.380799 to 0.41104, with only about a 2% gap remaining to the Meissner tetrahedron conjecture of 0.41986.

The official admitted that Hy4 preview is just an early iteration of Hy4, and there is still room for improvement in pre-training and post-training. It also has known issues such as complex task long thinking and excessive self-verification, and will continue to iterate agilely. The model has been deeply integrated with CodeBuddy and WorkBuddy, and users can also directly experience it in products like Yuanbao and ima.