Tencent Hunyuan Hy4 flash team to host Reddit AMA, claims 24x KV cache reduction

linjing219 · reddit · 2026-11-20

Tencent Hunyuan's Hy4 / Hy4 flash team (including Shunyu Yao) will host a Reddit AMA Friday 8–10 AM PT. Hy4 flash centers on a "12B active params + 1M context" combo; the team claims KV cache is cut to 1/6 of Hy4 and 1/24 of Hy3 via hybrid attention + LatentMoE + Engram — if it holds in real deployments, long-context high-concurrency inference cost changes by an order of magnitude. The author cautions the "1/20 per-task cost vs Hy4" figure is still a vendor claim with no third-party reproduction, and Artificial Analysis has no page yet. Three checkpoints shipped simultaneously (instruct, FP8, Base); Base scores MMLU-Pro 71.1, below Hy4 Base's 73.9, but C-Eval jumps to 92.0 — clearly tuned for Chinese office scenarios (WorkBuddy, docs, PPTs), leaving the general-capability trade-off an open question.

Related event: Tencent Hunyuan Hy4 Team to Host Reddit AMA(4 posts)→

Original post →

More from Models

Models channel →