Key Info

Tencent has released Hy4 Flash as an open-weight model, including Hy4 Flash-FP8 and Hy4 Flash-Base variants. It is a 224B-parameter MoE model with 12B active parameters per token, native image understanding, and 1M-token context.

Highlights

  • 224B total parameters with only 12B active per token, including a 64B Engram component
  • Natively multimodal with 1M-token context for documents, slides, and computer use
  • On Tencent's internal WorkIndex, it slightly outperforms the larger Hy4 Preview while cutting per-task cost to 1/13 of Hy4 Preview and 1/20 of Hy4
  • Decodes at over 150 tokens per second, with KV cache 1/24 the size of Hy3's and 1/6 of Hy4's
  • Architecture uses hybrid attention, LatentMoE, Engram, and a built-in DFly2 drafter; Day-0 vLLM support with a Docker image and deployment recipe