On August 3, 2026, MiniMax H3 was released with open weights, and ComfyUI shipped native support for it on the same day.
It's a next-generation model that handles text, images, and audio together, generating video with stereo audio — but there are some practical caveats that official blog posts and early reviews haven't covered much: the resolution ceiling for local runs, and regional restrictions in the commercial license.
This article covers setup on ComfyUI 0.30.0+, model selection by VRAM, how to use the three workflows (T2V/I2V/R2V), and licensing caveats, as of August 2026.
In my testing, the pruned INT8 diffusion model combined with the NVFP4 AWQ text encoder allowed MiniMax H3 to run on a GPU with 16GB of VRAM. Depending on your environment, you may still need to adjust settings such as resolution and video duration.
What is MiniMax H3?
MiniMax H3 is a video generation model released with open weights by China's MiniMax on August 3, 2026.
It's the third generation of the Hailuo series, and the first model the company has released weights for. ComfyUI added native support on release day.
Key specs:
- Model architecture: 33.1B-parameter single-stream omni transformer (about 13B of that is in AdaLN-modulation branches, which don't need to be loaded for inference-only use)
- Text encoder: Qwen3-VL-32B
- Output: up to 15 seconds, 24fps, with stereo audio
- Supported tasks: Text-to-Video (T2V), Image-to-Video (I2V, with start/end frame specification), Reference-to-Video (R2V)
- Supported languages: 11 languages, including Japanese
- Checkpoints: two families — FL2VA (handles T2V/I2V) and Ref2VA (handles reference-driven R2V)
Getting Started with MiniMax H3 in ComfyUI
Here's how to get MiniMax H3 running locally in ComfyUI, from setup through generation.
Step 1: Update ComfyUI to the latest version
Launch ComfyUI and update to the latest version before proceeding. Once the update finishes, move on to installing the model data.
Step 2: Get the model data
Next, download the model data needed for video generation.
The model is distributed via the Comfy-Org/MiniMax-H3 repository on Hugging Face. MiniMax's official repo (MiniMaxAI/MiniMax-H3) contains every precision and format, weighing in at around 498GB — but if you only download the files ComfyUI actually needs, it's more like a few dozen GB.
You need three pieces of model data for video generation:
- Video model data
- Text encoder
- VAE
Once you load the workflow, a list of required downloads appears. Download them in order, or use these commands to grab everything at once:
# Video model data
cd ComfyUI/models/diffusion_models
wget https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors
# Text encoder
cd ComfyUI/models/text_encoders
wget https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
# VAE
cd ComfyUI/models/vae
wget https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors
wget https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_fp16.safetensors
Step 3: Launch ComfyUI and open the workflow
This guide uses the official combined workflow.
Download the official workflow from Comfy's docs.
(Source: docs.comfy.org)
Drag and drop the downloaded workflow file into the ComfyUI canvas, then confirm the nodes have loaded correctly.
Step 4: Enter a prompt and start generation (T2V)
MiniMax H3's text-to-video mode can produce high-quality video from a fairly simple prompt.
Write your scene description in natural language in the prompt field, then click the "▷Queue" button to start generation.
Video generated with MiniMax H3 (ComfyUI):
Prompt used:
Emotional anime character film. The girl from Picture 1 in her original scene: a warm sunlit wooden hallway indoors, lit by dramatic warm golden backlight and soft ambient fill, deep soft shadow falloff into rich amber tones. Warm monochromatic palette with golden light and soft blue accents from her eyes. Emotional motif: a desperately reaching hand and a single falling tear. The environment is constant throughout.
SHOT 1: The scene opens exactly on Picture 1, the girl reaching desperately toward the camera; her eyes well with tears as the golden backlight pulses slightly brighter, dust motes drifting past as the camera executes a slow, deliberate push-in on her straining fingers and glistening eyes.
SHOT 2: Cut to an extreme close-up profile as her tear finally spills down her flushed cheek; the camera glides slowly alongside her face as loose strands of hair sweep across in the warm draft, her collar and necktie trembling faintly.
SHOT 3: Cut to a low, tilted wide shot: she loses her balance and tips forward, her twin tails sweeping wildly, the retreating figure's footstep fading further out of frame; the warm light flares gently along her damp eyes before the frame settles into a soft freeze as her hand falls just short.
Audio: soft ambient room tone, a trembling desperate breath, rustling fabric and a faint hair ornament chime, receding footsteps on wood, and a swelling emotional piano-and-strings score that resolves to a hushed, breathless quiet on the final beat.
Step 5: Generate video from an image (I2V)
MiniMax H3 can also generate video using a single reference image.
The workflow is available on the same official workflow page.
Structure your prompt by shot (SHOT 1–3), then click "▷Queue" to start generation.
Video generated with I2V:
Common Errors and Fixes
MiniMax H3 nodes don't show up in ComfyUI
Your ComfyUI version is older than 0.30.0. Update to the latest version. Note that Desktop and Cloud versions track stable releases, so the update may lag behind.
Running out of VRAM / high memory usage
Use pruned int8 instead of bf16 for the diffusion model, and nvfp4 AWQ for the text encoder. If it's still tight, lower the megapixel count in the Resolution Selector or shorten the duration.
Generation fails at low resolutions like 256p
MiniMax H3's minimum resolution is 384p. Anything at or below 256p will fail — pick from the resolution presets in the template instead.
No audio in the output video
Check that both minimax_h3_video_vae_fp16.safetensors (video) and minimax_h3_audio_vae_fp32.safetensors (audio) are loaded, and that the VAEDecodeAudio node is connected to the SaveVideo node.
Not sure which checkpoint to use for I2V
Use the FL2VA family for T2V and I2V (including start/end frame specification), and the Ref2VA family for reference-driven R2V. These are separate weights — the workflow won't function correctly if you mix them up.
MiniMax H3 Licensing and Commercial Use
MiniMax H3 is released under a custom license called the "MiniMax H3 Community License."
The main terms confirmed at the time of writing are as follows:
| Item | Details |
|---|---|
| Applicable Territory | Available worldwide, except for "Excluded Territories": the EU, UK, South Korea, and the US |
| Japan | Not listed as an Excluded Territory. Under the current license, both use and commercial use within Japan appear to be within scope |
| Commercial use | Permitted free of charge, but the "MiniMax H3" attribution must be displayed in the commercial product's UI |
| Revenue threshold | If annual revenue exceeds $20 million, separate written permission from MiniMax is required |
| Other | Using outputs to improve other AI models (e.g. distillation) is prohibited. Redistribution requires marking modified files and including a NOTICE file |
| Governing law | Hong Kong SAR law |
(Information as of August 2026.)
This regional-restriction clause has been a talking point on overseas technical blogs covering MiniMax H3.
If you're planning to integrate it into a commercial service, check the original license text (the LICENSE file on Hugging Face) with your legal team before making a decision.
Nothing in this article constitutes legal advice.
MiniMax H3/LICENSE: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE
Wrapping Up
This covered how to use MiniMax H3, the latest release in open video generation models.
It's one of the most capable open-weight video generation models currently available, and it's well worth exploring while it's still fresh.
This article was edited with AI assistance.
*Originally published in Japanese on EdgeHUB.
Top comments (0)