295B total / 21B active (MoE) + 3.8B MTP layer
256K tokens
Apache 2.0
Official post-preview Hy3 release from Tencent that follows Hy3-preview with scaled-up post-training, higher-quality data, stronger reinforcement learning, and public BF16 plus FP8 open-weight checkpoints. The official model card describes Hy3 as a 295B-parameter Mixture-of-Experts model with 21B active parameters, 3.8B MTP parameters, 192 experts with top-8 activation, 80 main layers, and 256K context. Tencent positions Hy3 as outperforming similar-size models and rivaling much larger flagship open models while improving practical utility in coding, office work, financial modeling, frontend design, game development, and broader productivity tasks. Release materials emphasize stronger agent capabilities, more reliable tool calling and output formats, lower hallucination and commonsense error rates, better long-context and multi-turn intent tracking, and more stable performance across different agent scaffoldings such as CodeBuddy, Cline, and KiloCode. Tencent also highlights a blind evaluation with 270 experts using work-derived tasks, where Hy3 reportedly scored 2.67 out of 4 versus 2.51 for GLM-5.1, plus companion deployment recipes for vLLM and SGLang with MTP-enabled speculative decoding and a released FP8 quantized checkpoint for more practical serving.