Tencent's Hy3 beats GLM-5.1 with 5x less compute
tencent released the hy3 moe model with 21b active params that reportedly outperforms larger models like glm-5.1 in expert evaluations.
tencent open-sourced hy3, a mixture-of-experts model with 295b total parameters but only 21b active per inference plus 3.8b for the mtp layer. it supports 256k token context and is available on hugging face under an apache 2.0 license, with an fp8-quantized version also released. in a blind evaluation of 270 experts, hy3 scored 2.67, beating glm-5.1's 2.51, while internal tests indicated a hallucination rate drop from 12.5% to 5.4%. integration has already begun across tencent's ecosystem, including wechat, workbuddy, and the path of exile game assistant.