Physix Frontier · News Briefing Card (QbitAI · Nov 21, 2026)
Tencent Hunyuan Reddit AMA: Hy small models and 24GB VRAM version in planning
KEY FACTS
- Yao Shunyu and the Tencent Hunyuan team held an AMA on Reddit about Hy4 and Hy4 flash.
- Hy4 was released on November 4, supporting text, image, and video understanding for long-horizon software engineering tasks.
- Hy4 flash was released on November 16, supporting a 1M token context with about 12 billion activated parameters per token.
- The team confirmed that Hy small models are already on the roadmap and plans to launch a model targeting 24GB VRAM.
- The team said it is adapting llama.cpp and collaborating with the vLLM and SGLang communities.
KEY DATA
1M tokensHy4 flash max context
about 12BHy4 flash activated parameters per token
about 156BHy4 flash total backbone parameters (excluding Eng
about 1/20Hy4 flash cost per task relative to Hy4
PHYSIX OBSERVATION
Hunyuan distills flagship capabilities into Flash and shortens the reasoning chain, seeking a balance between cost and intelligence. Small models and a 24GB VRAM version entering the roadmap shows that local deployment and the developer ecosystem are on the table; but llama.cpp adaptation has not yet landed, quantization accuracy still has pitfalls, and the speed of ecosystem catch-up will determine whether these models can truly be put to use.
Source: QbitAI report
Physix Frontier