Chinese AI company MiniMax has released the weights for H3-Image, a new image model that can generate a 4,096-by-4,096-pixel picture — roughly 16.8 megapixels — in about 10 seconds, without any upscaling step.
The release, posted to Hugging Face on Tuesday, makes the 33.1-billion-parameter model downloadable for anyone who wants to run it themselves. The same model is also available through MiniMax's API, its MiniMax Code tool and the Hailuo AI platform.
One model for generation and editing
H3-Image is a unified diffusion model, meaning a single checkpoint handles both text-to-image generation and image editing. It accepts up to 16 reference images in a single request, so users can feed it several pictures of a person, product or style and ask it to combine them.
MiniMax says the model reaches its result in just six sampling steps, which is the main reason it is fast. The company claims generation runs up to five times faster than mainstream alternatives, with most 4K images finishing in 10 to 20 seconds. It also supports wide aspect ratios, including 16:9 frames at 5,472 by 3,072 and 21:9 widescreen images at 6,272 by 2,688.
The model is built on MiniMax's H3 multimodal architecture, with the audio branch removed and a new image decoder added. It follows the company's H3 video model, which it open-sourced earlier, and arrives months after a MiniMax researcher said in a Reddit AMA in August that an image model based on the same architecture was in the works.
Open weights, but not open to everyone
The release comes with a notable catch. H3-Image ships under MiniMax's own community licence, not a standard permissive licence such as MIT or Apache 2.0, and that licence excludes users in the United States, the European Union, the United Kingdom and South Korea by default.
Businesses with annual revenue above $1 million, including affiliates, also need written permission before using the model commercially, as do companies offering hosted services to third parties. The licence even restricts where generated images can be displayed outside the permitted territory.
Running the model locally is also a heavy lift. The image-generation weights alone are about 68.65 GB, and the full setup, including the separate text encoder, totals roughly 135 to 138 GB — a machine with a high-end data-centre GPU such as Nvidia's H200 is the intended hardware, not a typical home PC.
For users who do not want to self-host, MiniMax prices H3-Image at about $0.069 per 4K image through its API, with a 50 percent discount during the first week. No benchmark scores accompanied the launch, so independent tests will decide how the model actually compares with closed rivals such as Google's image tools and OpenAI's image generation.
The release lands in a busy week for image models, and it underlines how quickly Chinese labs are moving to match — and in raw resolution, outpace — the closed models from the US. Whether developers outside the four excluded markets adopt it at scale will depend on how the licence restrictions play out in practice.
By Hamza Faisal



