Tencent’s Hy4 Preview: Pushing AI Boundaries

0
98

Tencent has officially unveiled Hy4 preview, a significant advancement in its Hunyuan large model series. This new iteration boasts a total of 770 billion parameters with 49 billion activated parameters and an impressive 1 million token context length. The company states that Hy4 preview solidifies its position at the forefront of open-source models, marked by substantial enhancements in model size, context window, and data scale.

Key Advancements and Applications

Hy4 preview has demonstrated remarkable progress across various real-world productivity tasks, thanks to collaborative data development with top experts from Tencent’s software engineering, gaming, finance, and security divisions. These advancements include:

  • Software Engineering: Enhanced understanding, planning, debugging, and verification capabilities for long-range development tasks, alongside improved visual aesthetics and interaction quality in front-end development.
  • Office Analysis: Significant improvements in comprehending complex office environments and financial analysis, with a focus on data analysis, cross-file collaboration, and completing the entire workflow from information processing to document, spreadsheet, and presentation delivery.
  • Game Development: The ability to generate playable prototypes directly from a single requirement, coupled with proficient use of game engines. Developers can refine complex game projects through multi-turn interactions.
  • Scientific Research: A notable boost in understanding, reasoning, and solving complex scientific problems. The model has made considerable progress in AI research and development, molecular dynamics simulations, condensed matter physics, and fundamental mathematics.

Performance Benchmarks

In a blind test involving 163 internal experts who evaluated 203 engineering tasks, Hy4 preview (average score 2.99/4.00) showed a slight edge over competing models. It performed marginally better than GLM-5.3 (average score 2.92/4.00), with a win rate of 46.8%, and Kimi K3 (average score 2.94/4.00), with a win rate of 51.2%.

Hy4 preview has also been instrumental in achieving a breakthrough on a classic geometric problem, the 3D Blaschke–Lebesgue problem. In conjunction with Hyra, it pushed the volume lower bound from 0.380799 to 0.41104. This significant advancement brings the solution very close to the final proof, with only a 2% gap remaining compared to the Meissner tetrahedron conjecture’s 0.41986.

Future Development and Availability

Tencent views Hy4 preview as an early iteration of the Hy4 model, acknowledging that both pre-training and post-training stages offer substantial room for improvement. Known issues, such as tendencies for lengthy thinking processes and excessive self-validation in complex tasks, are being actively addressed through agile development cycles.

“Similar to Hy3 preview, our goal in releasing Hy4 preview early is to gather extensive real-world feedback to significantly enhance the official Hy4 release. We will continue to leverage our unique advantage of deep collaboration with Tencent products and experts to continuously improve the inclusivity and ceiling of productivity.”

Hy4 preview is now open-source and available on Tencent Cloud TokenHub and OpenRouter. Users can also experience its capabilities through products like WorkBuddy, CodeBuddy, Yuanbao, and ima.

Links for access:

Source: https://www.ithome.com/0/995/570.htm

LEAVE A REPLY

Please enter your comment!
Please enter your name here