BREAKING
Tencent Open-Sources Hunyuan Hy3
0B
total params
0B
active params
0K
context length
SWE-bench Verified: 74.4%
Hy253
Hy374.4
Claude Opus 4.680.8
0%
faster TTFT
0%
faster response
0%
inference efficiency
Strengths and Shortfalls
PraisedPros
Runs locally via MoE
Strong cost-performance
Apache 2.0 open weights
CriticizedCons
Behind Qwen and GLM on hard tasks
Occasional switch to Chinese
Still trails top closed models
Tencent Pushes AI Comeback
AI NEWS BLITZ
Tencent has open-sourced its Hunyuan Hy3 foundation model under Apache 2.0.