Hunyuan Hy3

Tencent's flagship open MoE (Apache-2.0): 295B total / 21B active, 192 experts (top-8), 80 layers, 256K context, with a 3.8B multi-token-prediction layer for speculative decoding. Instruct- only (no separate base release). Tencent claims parity with GLM-5.2 and DeepSeek-V4 and a lead over GPT-5.5 on scientific tasks, with strong agentic performance. An FP8 sibling (Hy3-FP8) ships alongside. No AA Intelligence Index score at release.