Physix Frontier · News Briefing Card (IT Home · Nov 21, 2026)
Tencent Hunyuan Team Says Hy4 Has Room to Improve, Plans Larger Models
KEY FACTS
- Yao Shunyu and the Hunyuan team held a 120-minute AMA on Reddit, answering developer questions.
- The team said Hy3 and Hy4 prove that small models can outperform larger models at lower cost.
- The team said Hy4 still has room for improvement, is investing in large-scale reinforcement learning, and plans to train larger models.
- For dense models with 24 billion to 31 billion parameters, the team said it is possible and is under consideration.
- For coding-specific models under 100B, the team said it probably will not release one.
KEY DATA
120 minutesAMA duration
700 billionHy series parameter count
PHYSIX OBSERVATION
Hunyuan publicly admits a parameter gap with 3-trillion-scale rivals, yet chooses to build solid small models first before scaling up — a pragmatic route, but the risk lies in the time window. For developers, the near-term expectation is a better Hy4 experience from post-training, while ultra-large models and mid-to-small dense models still have no clear timeline.
Source: IT Home report
Physix Frontier