4K images and up to 16 references with H3-Image, MiniMax’s image model

H3-Image promises 4K images in ten seconds and supports up to 16 references. Its weights come with a license that restricts deployment in the EU.

The weights for H3-Image are available to download on Hugging Face, but permission to use them depends on location. MiniMax’s license excludes the European Union, the United Kingdom, South Korea, and the United States from its standard authorization. Organizations that want to deploy the model in those regions can apply for a separate license. Prior authorization is also required to offer a hosted service to third parties or when the annualized revenue of the user and its affiliates exceeds $1 million.

In a clarification published on Hugging Face, MiniMax says its API remains available globally, with safeguards enforced on its own infrastructure. The model is also available through MiniMax Code and Hailuo AI. Access therefore depends on whether users choose a hosted service or deploy the weights themselves. Generating and editing with the same model

The same set of weights handles both text-to-image generation and image editing. For edits, the API accepts between one and 16 reference images, which the prompt can address individually. In one of the examples published with the model, an empty room and nine furniture photos are combined into a furnished living room. Other examples turn a neutral 3D render into a photographic scene or change the viewpoint of a game map.

Prompt preparation is part of the workflow. By default, the API expands the supplied text into a detailed description and returns the version used for each image. Users can disable this expansion for text-to-image generation. According to the documentation, it always applies to editing requests. “4K” defined by image area

The “4K” tier targets an image area of roughly 16.8 MP, adjusted to the selected aspect ratio. The API documentation lists 4,096 × 4,096 for square images, 5,472 × 3,072 for 16:9 landscape, and 3,552 × 4,736 for 3:4 portrait. Custom dimensions are supported up to an area of 20 MP, within the permitted aspect ratios.

According to MiniMax, the model generates these large images directly. Generation takes six sampling steps. Files can be delivered as JPEG, PNG, or WebP. Transparent backgrounds and edits guided by a mask are among the features unavailable in the API at launch. A ten-second claim, with costs tied to each request

In its announcement, MiniMax claims it can generate a square 4,096 × 4,096 image in under ten seconds. Timing depends on the infrastructure and configuration. For example, the SGLang deployment guide reports about 85 seconds for a 4,736 × 3,552 generation on a DGX Spark machine, with the model already loaded and excluding network transfer and file saving.

The API offers Standard and Highspeed versions. The pricing page bills tokens for the text, reference images, and generated image. For a square 4,096 × 4,096 generation with no reference image, it gives example costs of about $0.069 in Standard and $0.130 in Highspeed, before the launch discount. Those amounts illustrate individual requests and vary with actual usage.

A 50% discount on the total request cost applies to delivered images of at least 16.6 MP until November 17 at 5 p.m. Paris time. During the promotion, the same Standard example costs about $0.034.