What actually changes

You add one file and drop the step count from 20 to 8. Nothing else in your setup moves. Measured here at 960x544 and 124 frames: 113 seconds becomes 53. At 4 steps it is 34 seconds, which is what I use while I am still hunting a prompt.

My rule: 4 steps while you are searching, 8 for the one you keep. When you do not yet know whether a prompt works, you do not need the best version of a bad idea. You need to find out quickly that it is a bad idea.

What you need

• minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors → ComfyUI/models/loras/ (this is the one that makes it fast)

• minimax_h3_fl2va_pruned_int8_convrot.safetensors → ComfyUI/models/diffusion_models/

• qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors → ComfyUI/models/text_encoders/

• minimax_h3_video_vae_fp16.safetensors → ComfyUI/models/vae/

• minimax_h3_audio_vae_fp32.safetensors → ComfyUI/models/vae/

The folder is the part people get wrong. A correct file in the wrong folder looks exactly like a missing model, and ComfyUI will not tell you which one it is.

Two things that still catch people

• The type box still has to say minimax. Changing the step count does not change that, and if it is wrong nothing runs.

• You need both decoders, one for the picture and one for the sound. Leaving either out fails in a way that does not obviously say so.

I wrote up every setting, what each one does and the mistakes that cost me time, in the full written guide. Free, and there is no card.

There is also a paid course that goes further with this model if you want it. Either way the workflow here is the whole thing, not a trimmed version.