Hy3 preview Free

Hy3 preview Free belongs to Tencent's Hy3 line of high-efficiency Mixture-of-Experts models, a family positioned for agentic workflows and production-grade reasoning tasks. Its architecture follows the broader Tencent MoE design, with a 295B-parameter backbone where roughly 21B parameters are active at inference, routed through 192 experts using a top-8 selection pattern. This configuration is intended to deliver strong reasoning and tool-use behavior while keeping per-token compute low, making the model well suited to interactive coding, retrieval-augmented agents, and long-context analytical workloads.

Within Tencent's Hy3 family, the preview variant supports configurable reasoning levels spanning disabled, low, and high modes, allowing developers to trade depth of inference against latency and cost depending on task complexity. Open-weight availability made it attractive for experimentation, while its large context window supported multi-document reasoning and extended code analysis. As the free-tier entry point of the family, Hy3 preview Free offered a zero-cost way to evaluate the architecture before moving to paid siblings, but it has since been marked deprecated, with active alternatives such as Hy3 Free and the full Hy3 release recommended for current deployments.