Screenshots show the official pages on the capture date and may differ from the pages at publication. Select an image to enlarge it.
On November 11, 2026, the Chinese AI company MiniMax released the image generation AI "MiniMax-H3-Image." A single model can generate images from text, generate images based on reference images, and edit images. The model's weights (its trained data) were released as open weights that anyone can download.
4K images in under 10 seconds
According to MiniMax, H3-Image can create 4K (4096×4096) images in under 10 seconds. The company says it is up to 5 times faster than major models from other companies, at 1.5 to 2 times faster in standard mode and 3 to 4 times faster in fast mode. The company says it gets this speed by cutting the image-making process down to as few as six steps, combined with improvements to its training data, training methods and serving system.
Its main features are as follows.
- It can create images of about 16 megapixels directly, without an upscaling step. Besides square images, it supports wide formats such as 21:9
- It can use up to 16 reference images
- MiniMax says it is the best open-weight model at drawing exactly what it is told
- In published samples, it correctly draws Chinese and English text inside images
The model is designed to be instructed with detailed descriptions in English. In the API, a built-in feature automatically rewrites short instructions in any language into detailed English descriptions.
Weights release and license
The core part that creates images has about 33.1 billion parameters, and a separate part that understands text is used alongside it. Its weights are published on Hugging Face, ModelScope and GitHub. However, the license is MiniMax's own "community license," not a permissive license like Apache or MIT. The license also includes provisions for cases that require separate written permission from MiniMax. Running it on your own machine requires about 69GB of data for the main model plus about 67GB for the part that understands text, so a high-performance GPU is a must.
Where you can use it and pricing
H3-Image is available on the video and image generation service "Hailuo AI," the MiniMax API and the development tool MiniMax Code. MiniMax says its prices are about 50% to 70% of those of major services from other companies. For the first week after launch, generating images at 4K or above costs half the usual price. Specific prices can be found on MiniMax's pricing page.
On the other hand, MiniMax has not published its evaluation scores as numbers.
CoAI's take
In image generation AI, paid services from US companies, such as Google's "Nano Banana" series, have led the way. What stands out about H3-Image is that a model with high-quality 4K output and fast speed has been released along with its weights. If you run it on your own servers, including the helper model that rewrites prompts, you can generate and edit images without sending images or prompts outside.
However, the license has its own conditions, so you need to check its terms before using it commercially. If you just want to try it out, the easiest way is to use it from the Hailuo AI interface.
