Image generation and editing in one model
Chinese AI company MiniMax on Nov. 11 announced the image generation model "MiniMax-H3-Image" on its official blog and released its weights (trained parameters) on Hugging Face. The model combines generating images from text and editing based on reference images into a single model. Up to 16 reference images can be used.
It is part of the company's "H3" series and uses a diffusion transformer (an AI approach that draws images out of noise) based on the H3 architecture with the audio layers removed. The company also has an official GitHub repository with the same description and license as on Hugging Face. Besides the released weights, the model is available through MiniMax's API and the company's apps "Hailuo AI" and "MiniMax Code."
MiniMax also released the weights of "M3.1 Flash," a model for language and AI agents, in October. The Hugging Face page for the new model was reported late on Nov. 10 Japan time by @aitrackerbot, an X account that tracks the appearance of models. The license text is also dated Nov. 10.
Company says 4K images in under 10 seconds
According to MiniMax, the model can produce a 4096 x 4096-pixel image (about 16 million pixels), which the company calls "4K," in under 10 seconds, up to 5 times faster than mainstream rival products. The company said it can render more detail than the 1 million to 4 million pixel outputs common today. Aspect ratios from 1:1 to 21:9 are available, and images of up to 6272 x 2688 pixels are produced directly, without a separate upscaling step.
The number of generation steps (sampling) was cut to 6. The blog said this was the result of combining data, reinforcement learning, fewer generation steps and deployment optimization, among other things. On the ability to follow instructions, the model card calls it the strongest among open-source models, and the blog says it is among the best of the open-weight models. However, neither the model card nor the blog names the rival products compared or gives numerical benchmark results.
Company says pricing is 50% to 70% of comparable mainstream services
The service offered through the API and elsewhere has 2 modes: the default "Standard" and the faster "Highspeed." MiniMax said that compared with mainstream rival products, Standard is 1.5 to 2 times faster and Highspeed 3 to 4 times faster. Pricing is 50% to 70% of comparable mainstream services, and during launch week, requests to generate images at 4K resolution or higher are 50% off the regular price. The blog does not give a specific per-image price.
About 69 gigabytes for the main model alone, plus a separate text reader
The released main model consists of a 33.1 billion parameter diffusion transformer and other components, with files totaling 68.65 gigabytes. The part that reads the instruction text (the text encoder) is not included; the model loads the one from MiniMax-H3, which is 66.7 gigabytes. The model assumes prompts that describe the whole image in detail in English. In the API, short prompts written in any language are automatically rewritten into detailed English before generation. The company also released "MiniMax-H3-Image-PE," a helper model that expands prompts when running the model locally.
It recommends SGLang, vLLM-Omni, diffusers and ComfyUI as frameworks to run it. The company said it runs on hardware such as NVIDIA's "H200" and "RTX PRO 6000 Blackwell" and AMD's "Instinct MI350X."
| Component | Role | Size |
|---|---|---|
| MiniMax-H3-Image | Image generation and editing (main model) | 68.65GB |
| Text encoder | Reads prompts (loaded from MiniMax-H3) | 66.7GB |
| MiniMax-H3-Image-PE (optional) | Rewrites short prompts into detailed English | 17.9GB |
License excludes U.S., EU, Britain and South Korea from its territory
The weights were released under the proprietary "MiniMax-H3-Image Community License Agreement." The grant is limited to the "Territory," which is defined as the world excluding the EU, Britain, South Korea and the U.S. The license does not permit use, reproduction or distribution outside the Territory. It asks those wishing to use the model in the excluded regions to contact MiniMax, which said it would grant licenses subject to strict controls and safeguards.
If the annualized revenue of the user and its affiliates exceeds $1 million, or if the model's capabilities are provided to third parties through an API, the web or other means, separate prior written permission from MiniMax is required. Permission is requested through an application form on the company's developer site. When used in commercial products or services, "MiniMax-H3-Image" must be prominently displayed on screen. When distributing to third parties, a copy of the license must be provided, along with a "NOTICE" file containing prescribed wording, except when offering the model as a service.
The license also prohibits using the model or the images it generates to improve other AI models, as well as use for military purposes. MiniMax said it does not claim rights over generated images. The helper model MiniMax-H3-Image-PE is based on "Qwen3.8-27B" (Apache 2.0 license), but the same license as the main model applies to it.
