comfy-minimax-modal

Use this skill when MiniMax H3 should run on Modal rather than on the local GPU. The workflow uses the repository's modal/minimax_h3.py app and keeps the result as an MP4 in the local workspace.

For MiniMax H3 generations, use the MiniMax-H3 Turbo LoRA and its four-step sampler by default. The preparation flow uploads loras/minimax/minimax_h3_turbo_4step_ckpt850.safetensors to the Volume and the Modal image installs ComfyUI-MiniMax-H3-Turbo. Pass --sage-attention by default; omit --easycache when audio quality matters. Only omit either accelerator when the user explicitly requests it or wants maximum quality/compatibility instead of speed. These flags are still opt-in in the CLI itself; this skill defines the agent's default behavior.

Bootstrap

If the modal command is missing, install it with: