Hy4 Flash is a cost-efficient multimodal Mixture-of-Experts model aimed at deep search, document processing, slide creation, and computer use. It carries 220B total parameters, including 64B of Engram n-gram embeddings, with 12B active per token across a 42-layer backbone of 3,072 hidden size. Twelve sparse attention layers interleave with 30 sliding-window layers, and LatentMoE routes each token to 16 of 512 experts plus a shared expert. Self-distilled from Hy4, it supports a 1M-token context with a SigLIP2-based vision encoder and a DFly2 drafter for speculative decoding.
Model specifications and capabilities are published by the model author and reproduced here from HF Model Card (tencent/Hy4-flash). Released under Tencent Hy License. Further reading: Blog
Choose your hardware and inference engine to get deployment commands and performance benchmarks tailored to your infrastructure
1,440 GB
Enter your email to get access to this content
Benchmarks measured by Vultr on B200 · vLLM. Throughput and latency vary with concurrency, input length and engine configuration, so treat these as a comparison baseline, not a service guarantee.