Tencent Hunyuan officially released the new open-source flagship large model Hy4preview on August 28. The model has a total of 770B parameters and 49B activated parameters, with a context length of 1M, targeting real productivity scenarios. The model has been simultaneously open-sourced on platforms such as HuggingFace, Github, and Modelscope, and is available on Tencent Cloud TokenHub and OpenRouter. Users can directly experience it in products such as WorkBuddy/CodeBuddy, Yuanbao, and ima.

QQ20260828-142733.png

In terms of capabilities, Hy4preview has made progress in four directions by leveraging high-quality data co-built with experts from Tencent's internal fields such as software engineering, gaming, finance, and security: enhancing long-term development tasks' understanding, planning, and debugging in software engineering; completing full delivery from data analysis to documents, tables, and presentations in office scenarios; supporting integration with Unreal5 and Unity engines via MCP for game development, generating playable demos from scratch through pure conversation; and covering areas such as molecular dynamics, condensed matter physics, and basic mathematics in scientific research.

QQ20260828-142743.png

Internal blind test data shows that among 203 engineering tasks scored by 163 experts, Hy4preview achieved an average score of 2.99/4.00, slightly better than GLM-5.3 (2.92) and Kimi K3 (2.94). In scientific research breakthroughs, combined with the Hyra model, it advanced the volume lower bound of the century-old classic problem of three-dimensional Blaschke–Lebesgue from 0.380799 to 0.41104, leaving only about 2% distance from the final proof of the Meissner tetrahedron conjecture; and achieved a 2.0x speedup in simulating a 32, 512-atom phospholipid bilayer, reaching 54.9ms/step.

Tencent stated that Hy4preview is an early version of the Hy4 iteration. Pre-training and post-training still have room for improvement, and there are known issues such as complex task long thinking. It will absorb real feedback through agile iteration. Its strategy of releasing an early version in an open-source manner continues the industry trend of shifting large model competition from parameter contests to productivity effectiveness verification.