Hy3 preview - The first step in rebuilding Tencent Hunyuan model
Hy3 Preview is a 295B-parameter MoE model (21B active) from Tencent. Built on a completely rebuilt RL infrastructure, it delivers best-in-class cost efficiency and significant improvements in complex reasoning, instruction following, and agentic coding tasks.
Replies
Flowtica Scribe
Hi everyone!
This is a milestone release from Tencent. After rebranding their Hunyuan foundation models to the "Hy" series, pronounced like "hi", this is the first model shipped since they rebuilt their pre-training and RL infrastructure in February 2026.
The model itself is a 295B MoE with 21B active parameters, 256K context, and three reasoning modes for different latency / depth tradeoffs. The biggest stated gains are in complex reasoning, context learning, instruction following, coding, and agent tasks, with support for @OpenClaw, @opencode, and @Kilo Code agent workflows.
It probably does not directly challenge the current frontier model landscape yet. But it does show something important: a very large ship is turning toward a much more aggressive direction. Definitely worth watching.
The model is already open-sourced, and you can try it for free on OpenRouter until May 8.