Hy3

Hy3 is a 295B-parameter Mixture-of-Experts (MoE) model with 21B active parameters and a 3.8B MTP layer, developed by the Tencent Hy Team. Following the Hy3 Preview launch in late April, the team scaled up post-training with higher-quality data and RL, gathering feedback from 50+ products. Hy3 outperforms similar-size models and rivals flagship open-source models with 2-5x the parameters, with strong gains in reasoning, agentic, and long-context tasks. It uses 80 layers (plus 1 MTP layer), 64 GQA attention heads (8 KV heads, head dim 128), a 4096 hidden size, 192 experts with top-8 activated, a 256K context window, and BF16 precision. Hy3 is a hybrid-thinking model supporting configurable reasoning effort (no_think, low, high), and emphasizes production-grade tool-call and output-format stability, reduced hallucination, and reliable multi-turn intent tracking.

Use the benchmark, Arena and pricing sections above as separate evidence. A missing field means the current data snapshot does not support that claim.