跳转到主要内容

Tencent Hunyuan Open-Source Flagship Model Hy4preview: 770B Parameters, 1M Context, Challenging Productivity Limits

Tencent Hunyuan dropped a bombshell on August 28 - the new open-source flagship model Hy4preview was officially unveiled. This model has a total of 770B parameters, 49B activated parameters, and a context length of up to 1M, clearly aiming to make a big impact in real productivity scenarios. Currently, Hy4preview has been open-sourced on platforms such as HuggingFace, Github, and Modelscope, and has been launched on Tencent Cloud TokenHub and OpenRouter. Users can also experience it directly in products like WorkBuddy/CodeBuddy, Yuanbao, and ima.

Image

Focusing on Four Directions, Targeting Real Needs

The highlight of Hy4preview is that it does not stop at the parameter competition, but goes deep into specific application scenarios. Experts from software engineering, gaming, finance, security, and other fields within Tencent jointly participated in building high-quality data, allowing the model to make substantial progress in four directions:

  • Software Engineering: Enhanced understanding, planning, and debugging capabilities for long-term development tasks, better able to handle complex programming challenges.
  • Office Scenarios: From data analysis to documents, spreadsheets, and presentations, Hy4preview can complete full-process delivery, taking office efficiency to the next level.
  • Game Development: Through MCP support, it integrates with Unreal5 and Unity engines, and can generate playable demos from scratch with pure conversation, opening new imaginative space for game developers.
  • Scientific Research: Covers molecular dynamics, condensed matter physics, basic mathematics, and other fields, aiding scientific breakthroughs.

Image

Hard Data Speaks, Strength Not to Be Underestimated

All talk and no action is fake. Tencent's internal blind test data shows that in the scoring of 203 engineering tasks by 163 experts, Hy4preview scored an average of 2.99/4.00, slightly outperforming GLM-5.3 (2.92) and Kimi K3 (2.94).

More excitingly, in the field of scientific research, Hy4preview combined with the Hyra model advanced the lower bound of the volume of the three-dimensional Blaschke–Lebesgue problem, a century-old problem, from 0.380799 to 0.41104, leaving only about 2% distance to the final proof of the Meissner tetrahedron conjecture. In addition, when simulating a 32,512-atom phospholipid bilayer, it achieved a 2.0x speedup, reaching 54.9ms/step.

Open Mindset, Embracing Feedback

Tencent admits that Hy4preview is only an early version of Hy4's iteration, with room for improvement in pre-training and post-training, and it is known to have issues such as long thinking on complex tasks. However, Tencent chose to release the early version in an open-source manner precisely to absorb real feedback through agile iteration. This strategy continues the industry trend - competition among large models is shifting from parameter comparison to validation of productivity efficiency.

Key Points

  • Tencent Hunyuan open-source flagship model Hy4preview, with total parameters of 770B, activated 49B, and context of 1M.
  • Breakthroughs in four major areas: software engineering, office work, game development, and scientific research.
  • Internal blind test score outperforms GLM-5.3 and Kimi K3.
  • Significant progress in the three-dimensional Blaschke–Lebesgue problem.
  • Tencent released the early version in an open-source manner, aiming for rapid iteration and feedback absorption.