MiniMax Opens the Weights of Its 16 Megapixel Image Model
MiniMax has published the weights for H3-Image, a text to image and image editing model built on its H3 family, and put them on Hugging Face for anyone inside its licence territory to download.
The company says a roughly 16 megapixel image, 4096 by 4096 at a square aspect ratio, takes under 10 seconds, which it describes as about five times faster than the norm for that resolution. Those are its own figures and no independent comparison was offered with the release.
And the design is economical. One checkpoint handles both generation and editing across up to 16 reference images, the diffusion transformer carries 33.1 billion parameters, it uses the H3 architecture without the audio layers, and it is distilled to six sampling steps without classifier free guidance, which is the part that explains the speed.
The packaging details matter to anyone planning to run it. The text encoder is a separate download, output sizes reach 6272 by 2688 without an upscaler, the BF16 checkpoint is 68.65GB, and an optional prompt expansion model adds another 17.9GB. MiniMax recommends SGLang, vLLM-Omni, diffusers and ComfyUI for local serving, with NVIDIA and AMD accelerators listed, and the model is also served through its own API and apps with half price requests at 16 megapixels and above for the first week.
MiniMax calls its instruction following the strongest among open image models and says the output avoids the familiar AI look, claims that will be tested by the people who download it rather than by the model card.
But the licence is the part worth reading twice. It is a community licence and not an open source one, it applies worldwide except the European Union, the United Kingdom, South Korea and the United States, and it says use outside that territory is not authorised by the agreement.
There are conditions inside the territory too, with separate written authorisation required from MiniMax for users whose annualised revenue passes USD 1 million or who host the model for third parties, and commercial products are asked to show the name MiniMax-H3-Image in their interface.
The release follows MiniMax's earlier open weights work on H3 for video and audio, so the company is building a family that developers can assemble rather than a single model.
So the useful way to read this is capability first and openness second, because a 33 billion parameter image model that runs in six steps and edits from 16 references is a genuinely useful download, and a licence that excludes four major markets is a commercial strategy wearing open weights as a costume.
MiniMax published open weights for H3-Image on Hugging Face
It reports 16 megapixel images in under 10 seconds at six sampling steps
The model edits with up to 16 reference images from a single checkpoint
The BF16 checkpoint is 68.65GB with a separate text encoder download
The community licence excludes the EU, UK, South Korea and the United States
Source: Pandaily
