Tencent releases open-source MoE model Hy3, claiming performance on par with models 5 times its size
Tencent releases Hy3 open-source model that allegedly matches models up to five times its active size
AI BriefingTencent has officially released its open-source MoE model Hy3, with 295B total parameters, 21B active parameters, an additional 3.8B MTP layer, and a 256k-token context window. Tencent says its performance rivals models two to five times its size. In a blind evaluation by 270 experts, Hy3 scored 2.67/4, surpassing GLM-5.1's 2.51; internal testing showed the hallucination rate dropping from 12.5% to 5.4%. Hy3 is available under the Apache 2.0 license on Hugging Face, ModelScope, and GitHub, with an FP8 quantized version and planned support for OpenRouter and Cline. Tencent has already integrated it into products including WorkBuddy, Yuanbao, WeChat, and the game Path of Exile: Advent.
Reported by 9 additional sources
- Tencent Hunyuan releases 1-bit and 4-bit versions of Hy3Tencent Hunyuan · July 15
- Tencent Hunyuan releases 1bit and 4bit quantized versions of Hy3, enabling the flagship AI model to run locally on a single GPUIT Home · Artificial Intelligence · July 14
- Tencent Hunyuan releases 1-bit/4-bit quantized versions of the Hy3 model; runs on a single GPUTencent Hunyuan · July 14
- Tencent Hunyuan releases quantized versions of Hy3: the 1-bit version can be deployed on a single GPU, while the 4-bit version delivers near-full performanceWeChat Official Account: Tencent Hunyuan · July 14
- Tencent Hunyuan Hy3 tops the OpenRouter rankingsTencent Hunyuan · July 13
- Tencent open-sources Hy3: a 295B MoE (21B active parameters); surpasses open-source models in search, trails GLM-5.2 in codingRohan Paul · July 7
- Tencent Hunyuan releases the 295B MoE model Hy3Tencent Hunyuan · July 6
- Tencent Yuanbao integrates the Hunyuan Hy3 model, supporting free file processingIT Home · Artificial Intelligence · July 6
- Tencent Hunyuan releases and open-sources the Hy3 model today, with performance comparable to flagship models 2–5 times its sizeIT Home · Artificial Intelligence · July 6
Source: The Decoder: AI News · the-decoder.com