MiniMax H3-Image: open weights, specs, settings and licence

What MiniMax released

The weights went up on Hugging Face alongside a post on MiniMax's blog, and the licence is dated 10 November 2026. H3-Image is MiniMax H3, the video model from August, with the audio branch taken out.

Generation and editing share one checkpoint. You can give it up to 16 reference images and call them <Picture 1>, <Picture 2> and so on in the prompt, numbered in the order you upload them. An edit is the same request with an instruction in place of a description.

There's also MiniMax-H3-Image-PE, an optional model for expanding short prompts. MiniMax's hosted version uses a different module for this, H3-Image-Context-IR, which also chooses the output size. Context-IR isn't in the open release.

Architecture and specs

With 32x32 pixels per token, a 4096x4096 image comes to 16,384 tokens. Many image models handle sequences of that length at 1 to 2 MP. MiniMax quotes under 10 seconds for a 4096x4096 image on its API. They don't name the hardware behind that number.

Resolutions and aspect ratios

MiniMax recommends staying near 16 MP and gives these sizes:

Smaller requests get rendered larger and scaled down. The reference server's min_generation_area defaults to 4,194,304 pixels (2048x2048). Ask for 1024x1024 and you get a 2048x2048 render resized with Lanczos. You can set it to 1024 to disable this. The model wasn't tuned for sizes that small.

Recommended settings

It's a distilled checkpoint. The SGLang and vLLM-Omni servers reject requests that change the fixed settings.

• Steps: 6. Anything else gets an HTTP 400.
• Guidance scale: 1, meaning no classifier-free guidance.
• Negative prompt: not supported; requests that include one are rejected. Describe what you want in the frame instead.
• Precision: BF16.
• Seed: set one for reproducible results; the card recommends creating the generator on the CPU.

import torchfrom diffusers import DiffusionPipelinepipe = DiffusionPipeline.from_pretrained(    "MiniMaxAI/MiniMax-H3-Image",    dtype=torch.bfloat16,    device_map="cuda",)image = pipe(    prompt="A detailed English description of the whole image...",    num_inference_steps=6,    generator=torch.Generator("cpu").manual_seed(0),).images[0]image.save("out.png")

Serving instructions cover SGLang (sglang serve --model-path MiniMaxAI/MiniMax-H3-Image) and vLLM-Omni. For lower latency you can split the sequence across several GPUs. Comfy has a tutorial with text-to-image and edit templates.

Download size and hardware

The text encoder isn't in the 68.65 GB H3-Image repo. You load it from the text_encoder/ folder of the MiniMax-H3 repo. Expect around 135 GB of downloads in total, plus 17.9 GB for the expander.

There's no official VRAM number. MiniMax lists the NVIDIA H200, RTX PRO 6000 Blackwell and AMD Instinct MI350X as supported. In BF16 the DiT is 66 GB and the text encoder 67 GB. Our guess is a 96 GB card with the text encoder offloaded, until quantised builds and measured numbers appear. Our hardware requirements page has the same kind of estimate for other models.

How to prompt it

According to the card, H3-Image "expects a detailed English description of the whole image." MiniMax's API gets away with short prompts because Context-IR rewrites them before generation. Locally you have two choices:

• Write the long description yourself, in English. Cover the subject, setting, framing, light and medium, plus any text you want in the image.
• Run a short prompt through H3-Image-PE (17.9 GB). It's built on a Qwen model and falls under the H3-Image licence.

Refer to reference images by number: "Put the jacket from <Picture 2> on the person in <Picture 1>, keep the background of <Picture 1>." Prompts over 4,096 tokens get truncated on the API.

Licence: what's allowed

The weights come under the MiniMax-H3-Image Community License Agreement. It's MiniMax's own licence, and the revenue limit is much lower than the H3 video licence's.

• Territory. No rights in the European Union, the United Kingdom, South Korea or the United States. MiniMax says it may change this and accepts licence requests from companies in those places.
• Revenue. Commercial use is allowed elsewhere. If you and your affiliates make more than US$1 million a year, you need written authorisation first. For H3 the limit is US$20 million.
• Hosting. Running it as a service for third parties needs written authorisation at any revenue.
• Attribution. Commercial products have to show "MiniMax-H3-Image" prominently in the UI. Anyone redistributing the weights must include the agreement and a NOTICE file.
• Training. You can't use the model or its outputs to improve other AI models. Derivatives of H3-Image are the exception.
• Outputs. MiniMax doesn't claim rights to what you generate.
• Acceptable use. The policy bans publishing generated content without disclosing it, impersonating people without consent, and automated high-stakes decisions such as credit, hiring or medical ones, among other things.

Can I use it on Windpaint?

No. Our customers include people in the excluded regions, and we don't have MiniMax's permission to host it. MiniMax H3 is listed as reference-only for the same reason. Models you can call through the API are on the image generation page.