Tencent Hunyuan Reddit AMA: mini and Flash-Lite models on roadmap, self-rated AGI level 2.5
腾讯混元 · wechat · 2026-11-21
After releasing and open-sourcing Hy4 and Hy4flash, the Tencent Hunyuan team held a Reddit AMA, fielding 300+ questions in two hours. Key takeaways:
Model roadmap
- Hy4/Hy4flash are general-purpose; mini models, 24–31B dense variants, and a frontier Flash-Lite (e.g. 35B-A3B) are on the roadmap; smaller open image models and a new video model are also coming
- Hy4 has lots of headroom via post-training; the team is heavily investing in large-scale RL, believes in Scaling Law, and plans significantly larger next-gen models
- Self-rated 2.5 on the Levels of AGI scale; RSI capability still early, with internal benchmarks under development
Architecture choices
- Hy4flash builds on iHC, gated MLA and DSA, adding hybrid SWA, Latent MoE and Engram; SWA cuts KV cache to 1/4
- Long GQA vs MLA tradeoff, resolved once DSA reduced MLA decode compute
- On MoE sparsity: more sparsity improves efficiency, but cutting active params means weaker models; overly sparse MoE is inference-unfriendly; edge models will get hardware co-design
- Quantization: down projections are more sensitive due to activation outliers; bit allocation by Hessian sensitivity; full-parameter QAT preferred, with router co-trained; 2-bit PTQ fails long before trit QAT
Training & safety
- Hardest post-training behavior to fix: overthinking; hardest moment: Hy3→Hy4 scaling with insufficient data ablations
- Biggest RL challenge: reward hacking from environment loopholes; tool-calling accuracy via hard cases and fine-grained rewards
- Cybersecurity training focuses on defensive tasks, excluding offensive exploit data; Hy4 is strong at vulnerability discovery but weak at exploitation
Related event: Tencent Hunyuan Team Details Hy4 and Flash Models in Reddit AMA(4 posts)→
More from Models
- Dev runs Gemini coding all day without hitting Pro limits; agent nails sound design — lipeng0820 · 2026-11-21
- Google appears to reset Ultra weekly limits for all Antigravity users — novice_invader · 2026-11-21
- Claims surface: Navier-Stokes and Riemann proofs via LEAN, by OpenAI and Claude — BasketVolleyball11 · 2026-11-21
- Yacine: 'It might be over for OpenAI — and the killer isn't Claude, it's Gemini' — yacineMTB · 2026-11-21
- InferenceBench ranks agents on LLM serving speedup: Claude Opus 5.5 tops at 12.08x, agents lose to simple hyperparameter search — maksym_andr · 2026-11-21
- Inkling-2-Mini quants are a trap: 17GB weights overflow 48GB VRAM at long context — EmilPi · 2026-11-21