MiniMax H3 Image released
- Source
- huggingface.co
- Author
- MiniMax model releases
- Date

Context
MiniMax has published the weights of MiniMax-H3-Image, a model that both creates images from a text prompt and edits existing ones, and its main technical claim is size at speed: it generates images of about 16 megapixels, such as 4096 by 4096 pixels, directly, without a separate upscaler. MiniMax notes that 1 to 4 megapixels is the common output today. The model is to six sampling steps without classifier-free guidance. A diffusion model builds an image by refining noise over many steps, and classifier-free guidance normally adds extra model passes at each step to make the result follow the prompt more closely. Fewer steps, each without that extra pass, mean less computation per image, which fits MiniMax's claim that 16-megapixel generation now comes at a practical cost; the card itself gives no timings and points to a separate deployment recipe for measured latency. The same checkpoint handles editing and accepts up to 16 reference images, and MiniMax claims the strongest instruction following among open-source models, though the model card shows no for that claim.
Running it locally is a large job. The image has 33.1 billion parameters and the repository holds about 69 GB, and the model also loads the 66.7 GB text encoder from MiniMax's earlier H3 model, which it reuses unchanged. MiniMax lists SGLang, vLLM-Omni, diffusers and ComfyUI as serving options and says both NVIDIA and AMD accelerators are supported, naming data-center parts such as NVIDIA's H200 and AMD's Instinct MI350X. The model expects a long, detailed English description of the whole image. MiniMax's API rewrites a short prompt in any language into that form, and a separate 17.9 GB prompt-expander model is offered for doing the same step locally.
The model card calls the release open source, but the license is narrower than that label suggests. It grants rights only outside its Excluded Territories, which it defines as the European Union, the United Kingdom, South Korea and the United States, and says use outside the permitted territory is not authorized. Elsewhere, a company whose annual revenue exceeds $1 million, or anyone offering the model as a hosted service to others, needs separate written permission from MiniMax. Developers and companies in those four regions therefore cannot rely on the downloadable weights under this license.
- distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
- benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
- transformer — The neural network architecture behind modern AI models, built on attention — letting every word directly consider every other word in parallel.
Checking sign-in…
Loading comments…


