Deploying MiniMax H3 Locally: Hardware Evaluation and ComfyUI Tutorial with RTX 5080 and 32GB RAM

> Windows Test Configuration: RTX 5080 16GB Video Memory, 32GB DDR4 Memory, ComfyUI + MiniMax H3 Director.

RTX 5080 16GB can run MiniMax H3, but it should use quantized weights; 32GB memory is sufficient, and it is recommended to set the page file to 48-64GB and reserve at least 60GB of disk space. It is recommended to start with 864×480, 24fps, 124 frames (about 5.17 seconds), 25 steps, res_multistep, simple, CFG 1.0, video shift 12, audio shift 3.

I. Weight Download

Official Repository: Comfy-Org/MiniMax-H3

• fl2va Text-to-Image/Video UNET, 20.97GB → models/diffusion_models/
• ref2va Reference Image/Video UNET, 20.97GB → models/diffusion_models/
• Qwen3-VL Text Encoder, 15.69GB → models/text_encoders/
• Video VAE, 4.85GB → models/vae/
• Audio VAE, 577MB → models/vae/

The first T2V requires fl2va, text encoder, and two VAEs; reference images/videos also require ref2va. Do not write Hugging Face tokens into the workflow or article.

II. Step-by-Step Deployment

1. Prepare Drivers and Directories. Install RTX 5080 drivers and run nvidia-smi in the terminal; it is recommended to use D:\minimax-h3\ComfyUI, and prepare models/diffusion_models, models/text_encoders, models/vae, and custom_nodes.

2. Start ComfyUI Once. Complete the installation using the official Windows installation method, then run:

cd D:\minimax-h3\ComfyUI
.\.venv\Scripts\python.exe main.py --listen 127.0.0.1 --port 8188

Browse to confirm the ComfyUI page appears then exit the program.

3. Install Director. Search for ComfyUI_MiniMaxH3_Director in ComfyUI Manager; or run:

cd D:\minimax-h3
git clone https://github.com/AIMixer/ComfyUI_MiniMaxH3_Director.git .\ComfyUI\custom_nodes\ComfyUI_MiniMaxH3_Director
.\ComfyUI\.venv\Scripts\python.exe -m pip install -r .\ComfyUI\custom_nodes\ComfyUI_MiniMaxH3_Director\requirements.txt

4. Add weights. Copy files to the target directory as described above, keeping the file names unchanged, and then restart ComfyUI.

5. Import workflow. Download and drag in the T2V workflow JSON. Select fl2va for UNET, Qwen3-VL for CLIP with the type set to minimax, and choose the video and audio VAEs separately.

6. Set parameters. Fill in the Director node with 864×480, 24 fps, 124 frames, 25 steps, res_multistep, simple, CFG 1.0, video shift 12, and audio shift 3. Connect CreateVideo and SaveVideo.

7. Complete the first test. Use the short prompt below, click Queue Prompt, and confirm that the MP4 contains both video and audio:

Realistic live-action cinematic shot of a quiet modern room at sunrise, soft natural light, subtle room ambience, no text, no watermark, no non-diegetic music.

8. Generate continuous stories. Each paragraph is controlled to about 5 seconds, with only one main action arranged. After generating the previous paragraph, extract the last frame as the first frame of the next paragraph's I2V, and reference <Picture 1> in the prompt words; continuously lock the character's clothing, weather, vehicle color, and left-right relationship. Subtitles are compressed in post-production, and BGM is uniformly added and faded between segments.

Resources

Official H3 Documentation　|　H3 Director GitHub　|　Chinese README