Skip to content
← Back to Videos

vLLM Adds Day-0 Hunyuan Hy4 Flash Support with 1M-Token Context

Loading video
Loading video

vLLM announced Day-0 support for Tencent Hunyuan Hy4 Flash, with an NVIDIA GPU integration built together with Inferact. The model targets multimodal research, document, slide and computer-use workloads with a 1M-token context. The announcement lists 224B total parameters including 64B of Engram and 12B active parameters, while the official deployment recipe lists 220B total. The recipe uses a 42-layer decoder with hybrid attention, Latent MoE, Engram and DFly speculative decoding, starting from an 8x H200, tensor-parallel-8 BF16 profile behind the vllm/vllm-openai:hy4-flash image.

Category: research
Author: @vllm_project
Date: 2026-11-18T00:00:00
Duration: 30.0s

As an Amazon Associate, we earn from qualifying purchases.