MiniMax releases open weights for H3-Image, a 4K image model that renders in under 10 seconds

Most image models output 1K or 2K images and rely on upscaling for anything larger. MiniMax says H3-Image generates native 4096 x 4096 (16 megapixel) images in under 10 seconds, at up to five times the speed of mainstream alternatives[1]. According to the model card, the diffusion transformer is distilled to six sampling steps with no classifier-free guidance, which is where much of that speed comes from[2].

The same checkpoint handles text-to-image generation and editing of one or more reference images, and MiniMax says it can work with up to 16 reference images[1]. The company claims state-of-the-art prompt adherence among open-weight models[1]. The speed and quality figures are MiniMax's own and have not been independently reproduced.

H3-Image is not a hobbyist download. Its diffusion transformer has 33.1 billion parameters stored in BF16, with a transformer folder of about 66 GB, and it loads the text encoder from MiniMax-H3, a further 66.7 GB if you do not already have it[2]. The model card includes serving instructions for SGLang and vLLM-Omni on a single GPU, or split across several GPUs for lower latency[2].

The weights ship under the MiniMax-H3-Image Community License, not a standard open source licence. It applies worldwide except the European Union, the United Kingdom, South Korea and the United States[3]. Inside the covered territories, which include Australia, organisations with more than US$1 million in annualised revenue, or anyone offering the model as a hosted service to third parties, must get prior written authorisation from MiniMax. Commercial products must also display "MiniMax-H3-Image" in their interface[3].

For teams that would rather not run it themselves, H3-Image is available through the MiniMax API Platform, MiniMax Code and Hailuo AI. MiniMax says its Standard mode runs 1.5 to 2 times as fast as mainstream alternatives, Highspeed mode 3 to 4 times as fast, and pricing sits at 50 to 70% of comparable offerings, with 50% off 4K and higher requests during launch week[1].

Open-weight image models have trailed proprietary ones on resolution and prompt adherence. A fast native 4K model with downloadable weights gives studios and developers in covered regions a self-hostable option, but the territorial carve-outs and revenue cap mean it is open in a narrower sense than Apache or MIT licensed releases.