Tencent/Hy3: 295B MoE model rivals trillion scale SOTA

AI & Machine Learning(huggingface.co)view on HackerNews
AImachine learningmodelMixture-of-ExpertsMoETencentHy3natural language processingNLP

Author: nateb2022

Date: 7/9/2026

Article Summary:
This article introduces Hy3, a 295B-parameter Mixture-of-Experts (MoE) model developed by Tencent Hy Team, which outperforms similar-size models and rivals flagship open-source models with 2-5x parameters.