MiniMax releases open-weight image model with native 16-megapixel output

MiniMax has released H3-Image, giving developers downloadable weights for generating and editing images at about 16 megapixels without a separate upscaling stage. The company says a 4096 × 4096 image takes under 10 seconds. Its release pairs that speed claim with a local deployment route and hosted access through MiniMax’s services.

The release builds on MiniMax’s H3 family, bringing text-to-image generation and reference-based editing into a single checkpoint—the trained model package. Users can supply up to 16 reference images. Rather than maintaining separate models for creating a picture and changing one, developers can use the same checkpoint for both tasks.

Six steps to a full-resolution image

H3-Image’s model card describes a 33.1-billion-parameter diffusion transformer, the component that predicts the image’s compressed representation. It uses six sampling steps, with no classifier-free guidance. The latter is an additional guidance technique that this model leaves out; its documented interface fixes the guidance setting at one and rejects negative prompts.

The model produces roughly the same pixel count across different image shapes. A square output is 4096 × 4096; an ultrawide 21:9 image is 6272 × 2688. MiniMax lists an effective working range of about four to 20 megapixels, with the best results generally around 16 megapixels—not a requirement that every image be square.

MiniMax attributes the release’s speed to changes across data, reinforcement learning, sampling-step reduction and deployment optimization, rather than one isolated improvement. The under-10-second result is a company performance claim, not a guaranteed completion time for every local installation.

Downloading the weights is only part of the setup

The downloadable package is substantial. MiniMax lists H3-Image at 68.65 GB and says its text encoder is a separate 66.7 GB download from the H3 repository. That encoder is not included in the image repository. These are file sizes, not a stated GPU-memory requirement, and an existing H3 installation can reuse the encoder.

For local serving, MiniMax recommends SGLang, vLLM-Omni, Diffusers, LightX2V and ComfyUI. Its instructions include single-GPU server commands and multi-GPU configurations that split a request across accelerators for lower latency. The weights are released under the MiniMax-H3-Image Community License Agreement, rather than an unrestricted grant implied by the word “open.”

Hosted access is available through the MiniMax API Platform, MiniMax Code and Hailuo AI. MiniMax says prices are 50–70% of comparable mainstream offerings. It also offers a launch-week discount of 50% off its standard price for requests at 4K resolution and above; that promotion is separate from its competitor-price comparison.

The prompt changes before the picture

A further distinction sits before image generation. H3-Image expects a detailed English description of the whole image. The hosted API expands short prompts in any language into detailed English descriptions before generation. Local operators can use MiniMax’s hosted expansion service or download the optional H3-Image-PE prompt-expansion model, listed at 17.9 GB.

MiniMax’s documented local workflow

• Expand the initial instruction into a detailed description and choose the output dimensions.
• Pass that description and size to the locally served image model.
• For editing, upload reference images in order and identify them in the prompt as Picture 1, Picture 2 and so on.