Newshot

Tencent Hunyuan's Yao Shunyu team: Hy4 still has significant room for improvement, plans to train larger-scale models

AI summary

At a Reddit AMA on November 21, Tencent's Hunyuan team said that Hy3 and Hy4 show well-built smaller models can outperform larger ones at lower cost, but Hy4 still has significant room for improvement, especially through post-training, and the team plans to train larger-scale models. The team also said a 24B–31B dense model is possible and under consideration, and that a small model benchmarked against Qwen 27B dense is "possible," though it likely won't release one.

Translation(model translation, not the source's own text)

Tencent Hunyuan's Yao Shunyu team: Hy4 still has much room for improvement, plans to train larger-scale models

IT之家, November 21: In the early hours of November 21, Yao Shunyu, Tencent Chief AI Scientist and head of the Hunyuan large model, and team members held a 120-minute AMA on Reddit, answering global developers' "anything" questions about Hunyuan.

During the Q&A, someone asked: ByteDance, Kimi, and others seem to be pushing toward ultra-large models with 3 trillion or even 5 trillion parameters, while you are still at 700 billion. Why? How do you view the actual impact of parameter count on model capability? What plans do you have in this area?

The Hunyuan team responded: "Hy3 and Hy4 have already shown that well-built small models can outperform larger models. They also cost less, and in the process, you can learn more and gain more."

However, the Hunyuan team also said that Hy4 still has a lot of room for improvement, especially through post-training. "We are investing a lot of effort in large-scale reinforcement learning, while continuing to improve user experience and model safety. We also plan to train larger-scale models."

As for whether it is possible to launch dense models with 24 billion to 31 billion parameters? The team said: "Possible, we are considering it."

As for whether they will release small models benchmarked against Qwen 27B dense models? Will they release models under 100B specifically for coding? The team said: "The first, possible. The second, probably not."

CDN media
CDN media

IT之家 note: Yao Shunyu graduated from Tsinghua University's Yao Class with a bachelor's degree and received a PhD in Computer Science from Princeton University. In 2024, he joined OpenAI and participated in the development of the agent products operator and deep research. His research focuses on language agents interacting with the real world; he proposed the ReAct method, pioneering the agent paradigm combining "reasoning—action," which has been widely adopted by academia and industry.

In December 2025, Tencent upgraded its large model R&D structure, newly establishing the AI Infra Department, AI Data Department, and Data Computing Platform Department, comprehensively strengthening its large model R&D system and core capabilities. Yao Shunyu became Chief AI Scientist of the "CEO / President's Office," reporting to Tencent President Liu Chiping; he also serves concurrently as head of the AI Infra Department and the Large Language Model Department, reporting to Lu Shan, President of the Technology and Engineering Group.

In July this year, Tencent announced the merger of the Hunyuan multimodal model department and the large language model department to establish the Foundation Model Department, managed uniformly by Yao Shunyu, to further improve model R&D and collaboration efficiency and explore the intelligence ceiling of omni-modal models.

Related articles

  • Tencent Hy4 officially open-sourced, Hunyuan LLM series fully open
  • Tencent: More enterprises turning to domestic chips, will help accelerate domestic supply chain construction
  • Tencent: Affected by chip restrictions, training next-generation models will take longer
  • Tencent Hunyuan Hy4 flash model officially announced as open source
  • Chinese AI flagship models collectively tighten cyberattack capabilities, strengthen AI technology export controls
  • Tencent Hunyuan Hy4 flash model released: 224B total parameters, 12B activated, cost per task about 5% of Hy4

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗