Minimax H3 Image

MiniMax's launch post for H3-Image, its open-weight model for generating images from text, generating from up to 16 reference images, and editing, makes speed at high resolution the headline: a 4096 by 4096 image, about 16 megapixels, in under 10 seconds. The post contrasts this with the more common 1K and 2K outputs, and says the model builds on its earlier H3 model with changes across post-training, including the training data, reinforcement learning and a reduction in sampling steps, the number of refinement passes needed to produce an image.

Every speed and price comparison in the post is against unnamed rivals. MiniMax says H3-Image reaches 4K at up to five times the speed of "mainstream alternatives", that its API's Standard mode is 1.5 to 2 times as fast and its Highspeed mode 3 to 4 times as fast, and that prices are 50 to 70 percent of comparable offerings. It does not say which models, hardware or settings it compared. Its claim of state-of-the-art prompt adherence among open-weight models also comes without a benchmark. Until independent tests appear, these figures are MiniMax's description of its own product. The model is offered through the MiniMax API Platform, MiniMax Code and Hailuo AI, with half-price 4K requests during launch week.

The "open weights" label needs a qualifier. The license on the Hugging Face release grants rights only outside its Excluded Territories, which are the European Union, the United Kingdom, South Korea and the United States. Even elsewhere, anyone whose annualized revenue exceeds $1 million, or who offers the model to others as a hosted service, needs separate written permission from MiniMax. Intel's explanation of the model card covers what running the weights takes, including the 33.1-billion-parameter image model and the large text encoder it reuses from H3.