MiniMax · 2026-11-11 · major
MiniMax H3-Image — open-weight model makes 4K images in under 10 seconds
In short
MiniMax H3-Image is an open-weight image model built on H3. It generates and edits 16 MP (4096×4096) images in 6 sampling steps, takes up to 16 reference images, and is #2 on Hugging Face trending.
MiniMax · 2026-11-11 · 3 sources checked
MiniMax's open-weight image model generates and edits 4K images in one checkpoint, in about 10 seconds.
Quick facts
| Maker | MiniMax |
|---|---|
| Model size | 33.1B-parameter diffusion transformer (68.65 GB BF16) |
| Output | About 16 MP, up to 4096×4096 (1:1) or 6272×2688 (21:9) |
| Sampling | Distilled to 6 steps, no classifier-free guidance |
| Reference images | Up to 16 per request |
| License | MiniMax-H3-Image Community License |
| Availability | Open weights on Hugging Face; MiniMax API, MiniMax Code, Hailuo AI |
What is it?
MiniMax H3-Image brings text-to-image generation, reference-based generation, and multi-image editing into one open-weight checkpoint. MiniMax built it on its H3 video model and says it generates a 4K (4096×4096) image in under 10 seconds. The weights are on Hugging Face under MiniMax's community license.
How does it work?
The core is a dense single-stream diffusion transformer with 50 layers that reads hidden states from the Qwen3-VL-based text encoder of MiniMax-H3; the audio layers of H3-Base are removed. A CNN-encoder, ViT-decoder VAE compresses each 32×32 pixel patch into one token, and sampling is distilled to 6 steps without classifier-free guidance. An optional prompt-expander model rewrites short prompts.
Why does it matter?
Most open image models stop well below 16 megapixels or need an upscaler for print-size output. MiniMax claims full-resolution 4K at up to 5× the speed of mainstream alternatives, and the model already runs in ComfyUI, SGLang and vLLM-Omni, so teams can self-host it instead of paying per image.
Who is it for?
designers, ad studios and teams self-hosting image generation
Frequently asked questions
- What hardware does MiniMax H3-Image need to run locally?
- MiniMax H3-Image needs about 135 GB in total when served with SGLang or vLLM-Omni, including the MiniMax-H3 text encoder. A single B200 or B300 holds the whole model, and a single H200 can run it by offloading the text encoder. The SGLang cookbook lists offloading recipes that go down to 12 GB cards.
- Is MiniMax H3-Image free for commercial use?
- MiniMax H3-Image ships under the MiniMax-H3-Image Community License Agreement, not a standard open-source license. Section IV of that license asks for a separate written authorization, which is requested through an application form on the MiniMax platform. The optional prompt-expander model, MiniMax-H3-Image-PE, falls under the same license.
- How much does the MiniMax H3-Image API cost?
- MiniMax says the H3-Image API costs 50–70% of comparable mainstream image offerings, without listing exact per-image prices in its launch post. For launch week, MiniMax offers 50% off the standard price for image requests at 4K resolution and above on the MiniMax API Platform.
- What can MiniMax H3-Image not do yet?
- MiniMax H3-Image does not yet support mask-based inpainting, transparent backgrounds, user-specified seeds, or preview images, according to its GitHub README. MiniMax also publishes no benchmark scores: its claim of the strongest instruction following among open-weight models comes without supporting numbers.
Try it
hf download MiniMaxAI/MiniMax-H3-Image