Tencent Open-Sources Hunyuan Hy3: 295B MoE, Hallucination Rate Drops to 5.4%

aigclink · x · 2026-07-06

Tencent has open-sourced the Hunyuan Hy3 large language model, featuring 295B total parameters and 21B activated parameters in an MoE architecture, with performance rivaling flagship models 2-5 times its size. It focuses on reliability in Agent scenarios: the hallucination rate dropped from 12.5% to 5.4%, common sense error rate from 25.4% to 12.7%, multi-turn dialogue failure rate from 17.4% to 7.9%, and long-context comprehension improved from 42.9% to 75.1%. It supports 256K context, switches between fast/slow thinking modes, and utilizes MTP speculative decoding to accelerate inference. In a blind test by 270 experts, Hy3 scored 2.67 (compared to GLM-5.1's 2.51), showing significant advantages in categories like frontend development, CI/CD, and data storage.

Related event: Tencent Open-Sources Hunyuan Hy3 for Agentic Workloads(26 posts)→

Original post →

More from coding & agent

coding & agent channel →