MiniMax releases H3-Image weights for 4K generation, with regional limits

MiniMax released H3-Image on November 10 UTC, making the weights of its 4K image generation and editing model available on Hugging Face. The company also announced access through its API, Hailuo AI and MiniMax Code. Developers evaluating a self-hosted setup face a geographic limit. The community license that accompanies the weights excludes South Korea, the United States, the United Kingdom and the European Union.

The agreement is dated November 10. It limits use, modification and redistribution of the model files to the territory it covers, and also restricts use, distribution and display of derivatives and generated outputs outside that territory. MiniMax says people in excluded regions who want to deploy the model can contact the company about a separate license.

Within the covered territory, a licensee needs prior written authorization if its annualized revenue or that of its affiliates exceeds $1 million, or if it offers the model to third parties through a hosted service. The agreement also requires the model name to appear prominently in a commercial product or service that uses it. MiniMax publishes separate specifications and pricing for its own hosted API.

Native 4K and up to 16 reference images

The model card lists a 4096-by-4096 square preset, or about 16.8 million pixels, and a 6272-by-2688 preset for a 21:9 frame. The model generates at those sizes directly, without an upscaling pass. The standard weights use six sampling steps, the iterations in which the model builds up an image.

The API can generate an image from text or take a prompt with up to 16 reference images for editing and reference-based generation. A product image and separate pictures of the material and background, for example, can inform the same composition. Developers can set width and height directly within the API's bounds. The maximum area is 20 million pixels, with aspect ratios from 1:3 to 3:1.

What self-hosting takes

The model card lists 68.65 GB of image model files and a separate 66.7 GB text encoder, the component that processes the prompt. Together the published file sizes come to about 135 GB. GPU memory usage during generation depends on which parts of the model are kept on the accelerator.

The vLLM team measured two ways to generate a 2048-by-2048 image on one Nvidia H200. With the text encoder offloaded, it recorded 73.4 GiB of peak GPU memory and a 3.2-second generation. Offloading the image generation component as well brought peak GPU memory down to 14.7 GiB and took 8.3 seconds. In a separate test with four H200s, the team reported a median end-to-end time of 3.84 seconds for a 4736-by-3552 image, based on three sequential requests to a warmed-up server.

What the two API tiers cost

MiniMax introduced H3-Image with a claim of 4K generation in under 10 seconds. Its published latency chart gives a median of 16.0 seconds on Standard and 10.3 seconds on Highspeed for a 4096-by-4096 text-to-image request. For a 4K edit with three reference images, the corresponding medians are 21.4 and 14.4 seconds.

MiniMax bills the API for tokens associated with text, reference images and the output. Its worked example for a square 4K image with no reference inputs comes to $0.06879 on Standard and $0.13001 on Highspeed before discounts. The launch promotion brings those examples to $0.03439 and $0.06501. Both tiers receive 50% off the request total when the delivered image contains at least 16.6 million pixels. The offer ends November 17 at 16:00 UTC.