Tencent Hunyuan team officially released the preview version of its new generation flagship model — Hy4preview today. The model has a total parameter count of 770B, with an activated parameter count of 49B, and supports a context length of up to 1M. With comprehensive expansion in model size, context length, and data scale, it demonstrates outstanding performance in real productivity tasks such as coding, office work, and scientific research, firmly placing it among the top open-source models. In multiple internal blind tests, its overall performance exceeded that of similar competitors.

image.png

In practical productivity scenarios, Hy4preview achieved multidimensional improvements through deep collaboration with experts from multiple fields within Tencent. In software engineering, it significantly enhanced long-range development understanding, planning, debugging, and verification capabilities, and optimized front-end visual aesthetics and interaction quality; in office analysis, it greatly improved the understanding of complex office environments and financial analysis capabilities, smoothly completing the entire delivery process from information processing to documents, tables, and presentations; in game development, it supports directly generating playable prototypes based on a single requirement and is proficient in using game engines for iterative development; in scientific research, it has made significant breakthroughs in AI R&D, molecular dynamics simulations, condensed matter physics, and fundamental mathematics. At the same time, it continues to deeply collaborate with tools like CodeBuddy and WorkBuddy, further refining the real user experience.

image.png

In the exploration of scientific frontiers and complex reasoning, Hy4preview also performed impressively. It can not only independently locate and optimize bottlenecks in the reasoning system, achieving a 31.8% improvement in end-to-end throughput compared to the baseline, but it can also coordinate multiple Codex Sessions like a researcher to organize experiments and continuously iterate. In machine learning force field molecular dynamics simulations, it worked with Hyra to achieve significant speedup, opening up more space for new material screening and drug development. In the design of low-temperature quantum transport devices in condensed matter physics, it independently completed the construction of a quantum scattering solver and the robust optimization of a five-barrier structure, significantly reducing the average leakage rate in high-energy bandgaps. In addition, it made significant progress on the classic geometric problem — the three-dimensional Blaschke–Lebesgue problem, pushing the volume lower bound up to 0.41104, leaving only 2% of the gap to the final proof of the conjecture.

Currently, Hy4preview is fully open-sourced and has been launched on Tencent Cloud TokenHub and OpenRouter, and users can also experience it through multiple products such as Yuanbao and ima. As an early preview version, the team will accelerate agile iteration based on extensive real feedback and continue to bring updates to the official version.