Run MiniMax H3 locally on an RTX 5090 or RTX PRO 6000

MiniMax H3 generates video with native stereo audio from text, keyframes or references. Here's the shortest path to running it in ComfyUI on one card.

Generator

33B · FL2VA (text/frames) + Ref2VA (references)

Text encoder

Qwen3-VL-32B

Output

4–15 s · 24 fps · stereo

Native size

1344 × 768

Pick your GPU

Choose once — every block below follows your pick.

Install ComfyUI

git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
python3 -m venv .venv && source .venv/bin/activate
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130
pip install -r requirements.txt "huggingface_hub[cli]"

Download the models

Files land straight in the right models/ folders.

hf download Comfy-Org/MiniMax-H3 \
  diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors \
  diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors \
  text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors \
  vae/minimax_h3_video_vae_int8_convrot.safetensors \
  vae/minimax_h3_audio_vae_fp32.safetensors \
  loras/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors \
  --local-dir models
hf download lightx2v/Minimax-h3-Turbo minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors --local-dir models/loras

Launch

python main.py

Open http://127.0.0.1:8188.

Generate

From text or a first frame: Workflow → Browse Templates → Video → MiniMax H3 T2V or I2V.

From your own references — images, a video clip and LoRAs — use this workflow. Drag the file onto the ComfyUI canvas:

• Images — upload into Reference image 1 / 2. Plug more Load Image nodes into the empty ref_image slots (up to 9).
• Video — upload a short clip into Reference video. No video? Select both video nodes and press Ctrl+B.
• LoRAs — drop files in models/loras/, pick one in Your LoRA 1 / 2, then Ctrl+B to switch it on.
• PRO 6000 — set Load Diffusion Model to minimax_h3_ref2va_bf16.safetensors.

Name references by tag in the prompt, in the order they're connected:

Use <Picture 1> as the main character — keep the face and outfit.
Use <Picture 2> for the location and colour palette.
Copy the camera move and pace from <Video 1>.
Shot: she walks toward camera, stops, looks up as the light shifts.
Audio: footsteps, room tone, one soft line of dialogue, no music.

My results

Same prompt and seed on both cards — filling this in as I test.

Sources

• MiniMax-H3 model card
• ComfyUI · MiniMax H3 native workflows
• Comfy-Org/MiniMax-H3 files