Alibaba's Wan team ships LoRAs that cut MiniMax-H3 video gen to 8 steps
Alibaba's Wan team applied Parallel Decoding Distillation to MiniMax-H3 and released official acceleration LoRAs for both the FL2VA and Ref2VA checkpoints. Each rank-64 BF16 LoRA lets the video model generate in as few as 4–8 neural function evaluations, so you swap in a LoRA instead of running a whole new model. Load it with Diffusers 0.40+ via the provided predict scripts.