Tencent Hy3

Hy3 is a 295B-parameter Mixture-of-Experts language model from Tencent (21B active parameters, 192 experts with top-8 routing, plus a 3.8B MTP layer), developed by the Tencent Hy Team and released under Apache 2.0. Refined on the Hy3 preview release with improved post-training data and scaled RL, it targets reasoning, agentic workflows, coding and long-context tasks, supporting a 256K context window, hybrid fast/slow thinking via a configurable reasoning_effort parameter, function calling, structured output and caching. It is served on Tencent Cloud TokenHub with an OpenAI-compatible API.