Tencent Ships Hy3, a 295B Apache-2.0 MoE That Matches Flagship Open Models at a Fraction of the Active Parameters

Summary

Tencent has released Hy3, an open-weight Mixture-of-Experts language model with 295B total parameters but only ~21B activated per token (plus a 3.8B MTP layer), under the permissive Apache 2.0 license. Tencent claims Hy3 "outperforms similar-size models and rivals flagship open-source models with 2–5x parameters," supports a 256K-token context window, and ships as a 598GB full model plus a 300GB FP8 quantized variant on Hugging Face. Free access is being offered via OpenRouter through July 21, 2