跳转到主要内容

Tencent Hunyuan Large Model Upgrade: 770B Flagship New Version Hy4preview Open-Sourced, Directly Entering the Top Tier of Open-Source Models

Tencent Hunyuan team today officially released the preview version of its new flagship model - Hy4preview. This model has a total parameter count of 770B, activated parameters of 49B, and supports a maximum context length of 1M. After comprehensive expansion in model scale, context length, and data scale, it has demonstrated excellent performance in real productivity tasks such as programming, office work, and scientific research, firmly ranking in the top tier of open-source models. In multiple internal blind tests, its overall performance even surpassed similar competitors.

Image

In real productivity scenarios, Hy4preview has achieved multi-dimensional capability improvements through deep collaboration with internal experts from various fields at Tencent. In software engineering, it significantly enhances long-distance development understanding, planning, debugging, and verification capabilities, and optimizes front-end visual aesthetics and interaction quality; in office analysis, it greatly improves understanding of complex office environments and financial analysis, smoothly completing the entire delivery process from information processing to documents, spreadsheets, and presentations; in game development, it supports generating playable prototypes directly from a single requirement and can skillfully use game engines for iterative development; in scientific research, it has made significant breakthroughs in AI research, molecular dynamics simulation, condensed matter physics, and basic mathematics. At the same time, it continues to deeply cooperate with tools such as CodeBuddy and WorkBuddy to further polish the real user experience.

Image

In scientific frontier exploration and complex reasoning, Hy4preview also performs outstandingly. It can not only autonomously locate and optimize bottlenecks in reasoning systems, achieving an end-to-end throughput improvement of 31.8% compared to the baseline, but also coordinate multiple Codex Sessions like a researcher to organize experiments and iterate continuously. In machine learning force field molecular dynamics simulations, it works collaboratively with Hyra to achieve significant acceleration, 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 robust optimization of a five-barrier structure, significantly reducing the average leakage rate in high-energy band gaps. In addition, it has made significant progress on the classic geometric problem - the three-dimensional Blaschke–Lebesgue problem, pushing the volume lower bound to 0.41104, leaving only a 2% gap to the final proof of the conjecture.

Currently, Hy4preview has been fully open-sourced and is available on Tencent Cloud TokenHub and OpenRouter. 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 a large amount of real feedback and continue to bring updates to the official version.

Key Points

  • Model Scale: Total parameters 770B, activated parameters 49B, supports 1M context.
  • Performance: Excellent in programming, office work, game development, scientific research, etc., surpassing competitors in multiple blind tests.
  • Open Source and Availability: Fully open-sourced, available on Tencent Cloud TokenHub and OpenRouter, can be experienced through products such as Yuanbao and ima.
  • Future Plans: The team will accelerate iteration based on feedback, and the official version will be continuously updated.