Hasty Briefsbeta

Bilingual

Tencent Hy3

7 hours ago
  • #Mixture-of-Experts
  • #Deployment Guide
  • #AI Model
  • The text provides instructions on how to use the tencent/Hy3 model with various libraries, inference providers, notebooks, and local apps.
  • Detailed examples are given for using Hy3 with Transformers, vLLM, SGLang, and Docker, including code snippets for installation, server setup, and API calls.
  • Hy3 is a 295B-parameter Mixture-of-Experts (MoE) model with 21B active parameters, developed by Tencent Hy Team, offering strong capabilities in agent tasks, reasoning, and long-context handling.
  • The model shows improvements in stability, knowledge accuracy, and multi-turn intent tracking, with reduced hallucination and error rates based on internal evaluations.
  • Deployment recommendations include using vLLM or SGLang on 8 GPUs with specific configurations, and the model is open-sourced under Apache License 2.0, available on multiple platforms.