Tencent Releases Hy3: An Open 295B Mixture-of-Experts (MoE) Model with 21B Active Parameters and 256K Context

Hy3 is an open-source language model with 295 billion total parameters that only activates 21 billion parameters per token, using a sparse mixture-of-experts architecture designed for reasoning tasks, agent workflows, and processing long documents up to 256,000 tokens. The model incorporates a multi-token prediction layer for faster decoding and includes training focused on production reliability, including improvements to tool calling stability, reduction of hallucinations, and better tracking of multi-turn conversations. Performance benchmarks show Hy3 scores 78.0 on coding tests and 90.4 on reasoning tasks, with blind testing indicating advantages over competing models in frontend development and infrastructure work. The weights are released under the Apache License 2.0 and can be deployed using vLLM or SGLang, with a free trial available through OpenRouter.