Hy3 Preview

Hy3 Preview uses a MoE architecture with 295B total parameters and 21B active per inference.

Context length is 256K (input 192K / output 128K), holding roughly 140,000 Chinese characters; it supports interleaved deep thinking, JSON Schema structured output, Function Calling, and cache.

Use cases include long documents, Agent orchestration, coding assistance, knowledge-base Q&A, and contract review—MoE activation keeps serving cost practical while delivering high performance on long-context and tool-heavy workloads.