Image: MiniMax
MiniMax released H3-Image, an image model that generates and edits pictures at roughly 16 megapixels, and put the weights on Hugging Face. The company's pitch is "4K image generation in 10 seconds." That number is doing some work, so it is worth unpacking before you plan anything around it.
Sixteen megapixels means 4096 by 4096 pixels. The model holds that figure across aspect ratios, from square up to 6272 by 2688 for a 21:9 banner. MiniMax says it renders at that size directly, without a separate upscaler pass. For scale, a 4K video frame is 3840 by 2160, about 8.3 megapixels. So one H3-Image picture carries roughly twice the pixels of a 4K frame and about eight times a 1080p one. It is a print and crop-resolution claim, not a video one.
The speed comes from a specific decision. The model is distilled to six sampling steps with no classifier-free guidance. In plain terms, it needs six passes over the image instead of the 30 to 50 that diffusion models typically take, and it skips the guidance trick that doubles the work at every step. This is not a setting you can tune. The API returns an error if you ask for any number of steps other than six, or any guidance scale other than one, and it rejects negative prompts outright.
4K image generation in 10 seconds. Meet MiniMax H3-Image. 16 million pixel resolution. Stunning realism. Precise prompt following and editing. Up to 16 reference images. All in 10 seconds. And it's open-weight. Available now via our API, MiniMax Code, and Hailuo AI, with 50% off 16MP+ generations during launch week.
— MiniMax (@MiniMax_AI) November 10, 2026
Open weight does not mean light. The diffusion transformer is 68.65GB in BF16, and it borrows a 66.7GB text encoder from MiniMax-H3, plus a 2.34GB image VAE. That is about 138GB of files before you add anything else. MiniMax points at SGLang, vLLM-Omni, diffusers and ComfyUI for local serving, and lists H200, RTX PRO 6000 Blackwell and AMD MI350X among the supported accelerators. If you do not have that hardware, the API is the realistic path.
The licence is a community agreement, not Apache 2.0 or MIT. Commercial use has conditions attached, and the card points to a separate written authorisation process through an application form for some uses. Read it before you build a product on top of the weights.
Generation and editing share one checkpoint, which is the tidier design choice here. Editing accepts up to 16 reference images, and their upload order becomes the <Picture 1>, <Picture 2> labels you refer to in the prompt. The card's own edit example removes headphones from a podcast photo and has the two hosts bump fists. Another composes nine pieces of furniture into an empty room from ten references.
The detail most people will trip over is prompting. H3-Image expects a detailed English description of the whole picture, not a short phrase. The MiniMax API quietly rewrites your prompt into one using a component called H3-Image-Context-IR before generating, and there is a local 17.9GB prompt expander if you want to stay offline. Skip that step and you are feeding the model the wrong shape of input, then blaming the model for the result.
There is a second quiet quirk in the size handling. The API has a minimum generation area that defaults to 2048 by 2048. Request anything smaller and it renders at your aspect ratio on a larger canvas, then downsamples with Lanczos. That is why a small output can look slightly soft. Passing 1024 turns the behaviour off.
MiniMax claims the strongest instruction following among open image models. That is a company claim and no third party has published a comparison yet, though the image leaderboards have been busy lately: Microsoft's MAI-Image-2.7 Preview moved up Arena's image-editing board last month. It is also one of the few recent image releases to ship open weights at all, which puts it in the same small group as Baidu's ERNIE 5.5 Flash Base.
H3-Image is available now through the MiniMax API, MiniMax Code and Hailuo AI, with half off 16MP and larger generations during launch week. For a design studio that needs large source files, or a team that wants to fine-tune an image model in-house, this is the first open option in a while worth a serious test. For everyone else, the honest summary is simpler. It is fast, it is large, and the open weights are real but heavy.
Comments