MiniMax H3 Image in ComfyUI: file sizes and licence catch
MiniMax H3 Image runs in ComfyUI 0.45.0. What the local files weigh, what the templates need, and why its licence excludes the US, EU, UK and Korea.
· 3 min read
MiniMax H3 Image, an open-weight model for text-to-image generation and image editing, runs locally in ComfyUI from version 0.45.0, released on 10 November 2026. Before you download about 49 GB of files, read the licence: it grants no rights in the United States, the European Union, the United Kingdom or South Korea.
This brief covers what was released, what the ComfyUI files weigh, and what the licence allows. It is for anyone who runs ComfyUI on their own graphics card.
What happened
MiniMax published MiniMax-H3-Image on Hugging Face with a licence dated 10 November 2026. One checkpoint handles both text-to-image and editing with up to 16 reference images. The model card describes a 33.1B-parameter diffusion transformer, distilled to six sampling steps without classifier-free guidance, with native output of about 16 megapixels (4096×4096 at 1:1, up to 6272×2688 at 21:9).
The same day, ComfyUI v0.45.0 added "Support MiniMax-H3-image", and Comfy's announcement of 10 November added text-to-image and edit templates. The ComfyUI tutorial asks for ComfyUI 0.45.0 or later and template pack 0.11.104 or later. It also says to keep the templates' six steps at guidance 1, with no negative prompt.
How much does MiniMax H3 Image need to run locally?
Neither MiniMax nor Comfy publishes a VRAM figure. The model card names NVIDIA H200, RTX PRO 6000 Blackwell and AMD Instinct MI350X among the supported accelerators. It does not mention Apple silicon. Download sizes are known, though. These are the files in Comfy-Org's repackaged repository, from its file listing on 12 November 2026:
| File | Precision | Size |
|---|---|---|
| Diffusion model (template default) | INT8 | 20.98 GB |
| Diffusion model | W4A8 | 12.55 GB |
| Diffusion model | BF16 | 40.23 GB |
| Qwen3-VL 32B text encoder (template default) | INT8 | 27.14 GB |
| Qwen3-VL 32B text encoder | NVFP4 | 15.69 GB |
| Qwen3.5 9B prompt expander (optional) | W4A8 | 7.34 GB |
| VAE | FP16 | 1.17 GB |
The tutorial's default set comes to about 49.3 GB, or 56.6 GB with the prompt expander. Smaller variants are in the repository, but the tutorial does not say which hardware each precision is meant for. The text-to-image template starts at 4 megapixels "to suit low-VRAM machines". If an edit runs out of memory, the tutorial suggests lowering the resolution or turning on the Resize Image/Mask node.
Where does the MiniMax H3 Image licence apply?
Inside that territory, the same licence requires prior written authorisation from MiniMax if you and your affiliates have annualised revenue over 1 million US dollars, or if you host the model as a service for others. It also forbids using the model or its outputs to improve other AI models, and it requires machine-generated content you post publicly to be disclosed as such. MiniMax says people in the excluded territories can contact it about a licence. This is a summary, not legal advice.
If you are outside those four territories, the weights and ComfyUI's templates are ready to use today. If you are inside one, anyone can download the weights, but the licence does not give you the right to use them.
Using it from your other devices
Corewire can forward ComfyUI's web interface from the machine with the graphics card to your laptop over an encrypted connection. The forward is off until you switch it on at that machine. The setup is in the ComfyUI guide, and the app is on the download page. If you are choosing hardware for language models rather than images, see best local LLM by memory.