MiniMax open-sources H3-Image, 33.1B-parameter diffusion model
TL;DR
- MiniMax has open-sourced H3-Image, a unified diffusion model that handles text-to-image and image editing from a single 33.1B-parameter checkpoint.
- The model targets ~16 megapixel output at resolutions up to 6272x2688, generated directly without a separate upscaling step.
- Editing accepts up to 16 reference images; inference runs in six sampling steps with no classifier-free guidance.
MiniMax has posted MiniMax-H3-Image, described on the Hugging Face card as a "unified diffusion model that performs both text-to-image generation and image editing from a single checkpoint." The DiT backbone is listed at "33.1B parameters, 50 layers, dense single-stream transformer with 3D multimodal RoPE," paired with a full Qwen3-VL text encoder that reads its 50th layer hidden states.
The headline is resolution. The card says the model generates "~16 megapixels at any aspect ratio (4096x4096 to 6272x2688)" directly, without an upscaler. Editing "supports up to 16 reference images." Inference runs in "6 sampling steps (6 NFEs) with no classifier-free guidance required."
The checkpoint is not light. The DiT weights are 66.28 GB on disk, with the separate text encoder adding another 66.7 GB. Two of the researchers we follow posted the Hugging Face link the day it went up.
Shared on Bluesky by 2 AI experts
Originally reported by huggingface.co
Read the original article →Original headline: MiniMaxAI/MiniMax-H3-Image · Hugging Face