MiniMax releases H3-Image, an open-weight model for native 4K image generation and editing

MiniMax released H3-Image, an image generation and editing model with open weights on Hugging Face. MiniMax says it renders 16-megapixel (4K) images in about 10 seconds, follows prompts and edits precisely, and accepts up to 16 reference images. These are MiniMax's claims.

It is available through MiniMax's API, MiniMax Code and Hailuo AI, with 50% off 16MP-and-larger generations during launch week.

The SGLang project says SGLang Diffusion supported the model from day 0 at full precision, with multi-GPU inference and recipes for DGX Spark, RTX PRO 6000, H200 and B300. By its own measurements, a 4K image takes 85 seconds on a DGX Spark, 18.5 seconds on one RTX PRO 6000 and 10 seconds on one H200 (2K: 16, 3.4 and 1.9 seconds). It says precomputing AdaLN saves 24.6 GiB of VRAM.