MiniMax has released the weights for H3-Image, but anyone who wants to run it on their own machine hits the first obstacle in the download size. Going by the files listed on the Hugging Face model card, the DiT diffusion backbone is about 68.65GB, the text encoder about 66.7GB and the image VAE 2.34GB, for a total of roughly 138GB, before counting the optional prompt-expansion model.
The text encoder is not in the H3-Image repository; it has to be downloaded separately from the repository of the H3 video model. Teams that have already deployed the H3 video model can reuse that part and cut the download roughly in half.
What hardware can run it
The model card gives no VRAM floor. It only lists the platforms the inference frameworks have been verified on: NVIDIA H200, RTX PRO 6000 Blackwell and AMD Instinct MI350X. Single-card operation is supported, and four-card sequence parallelism can be used to cut latency.
The three cards carry 141GB (H200), 96GB (RTX PRO 6000) and 288GB (MI350X) of memory respectively. The BF16 weights alone exceed 100GB, so a single card will most likely have to rely on staged loading or moving the text encoder elsewhere; the exact configuration should follow the vLLM-Omni and SGLang deployment recipes the official side provides. Consumer GPUs, including 24GB and 32GB models, currently appear to fall outside official support.
The optional prompt-expansion model, H3-Image-PE, has a much lower bar. It is fine-tuned from Qwen3.8-27B and released as NVFP4-quantized weights of about 17.9GB. Run locally with Diffusers it needs roughly 19GiB of VRAM, so it fits on a single 24GB card. Its job is to expand a one-line request into a full scene description and recommend an aspect ratio. It can be deployed on its own, with its output handed to the cloud-hosted H3-Image for the actual drawing.
How developers in China can get it
The weights are available on both Hugging Face and ModelScope, and downloading from ModelScope within China is considerably smoother. Four deployment routes are offered: SGLang, vLLM-Omni, Diffusers and ComfyUI. Designers used to node-based workflows will likely find ComfyUI the quickest way in once its node package stabilizes, while teams building services are more likely to pick SGLang or vLLM-Omni.
The license is the MiniMax-H3-Image Community License Agreement, which also applies to the PE model. Uses covered by Section 4 of the agreement require written authorization through an official form, so check which category you fall into before any commercial use. The model card does not spell out separate regional restrictions. A third-party breakdown earlier noted that the community license for the H3 video model excludes certain countries and regions; whether the H3-Image agreement carries that over should be checked clause by clause before downloading.
Options if you have no GPU
For most individual creators and small teams, deploying a 138GB model on their own hardware is not a good deal. Official channels include the MiniMax Open Platform API, Hailuo and MiniMax Agent; if you would rather not integrate an API, you can use the model directly on partner platforms.
PixPix focuses on e-commerce and product images, with ready-made entry points for product image sets, detail pages and model shots; choose H3-Image there to get 4K output. PixTV is an infinite canvas on which image, video and audio models can all be called, which suits a workflow of locking in a still frame with H3-Image and then handing it to a video model for the final cut. Both are pay-per-use and need no hardware on your side.
A rough line between self-hosting and calling a service might look like this: if you generate a few hundred images a day or fewer and have no hard requirement that data stay local, use the web interface or an API; if your volume is large, you need to modify the model, or you have private-deployment requirements, then consider putting GPUs behind it.
Sources: Hugging Face model card, CocoLoop, MiniMax-H3-Image-PE model card, MiniMax official blog; component file sizes, supported hardware platforms, deployment frameworks and the license agreement name were checked.