ModelsNews
MiniMax H3-Image: open 16 MP image model, minus the EU
MiniMax released H3-Image, an open-weight image generation and editing model. It outputs about 16 megapixels in 6 steps without upscaling and takes up to 16 reference images. The license excludes the EU, the UK, South Korea and the US.
· 2 min read

What MiniMax released
MiniMax-H3-Image arrived on November 10, derived from the MiniMax H3 video model family. A single checkpoint handles text-to-image generation and editing of reference images (Hugging Face).
- Resolution: about 16 megapixels directly, from 4096x4096 (1:1) to 6272x2688 (21:9), with no upscaler.
- Speed: distilled to 6 sampling steps with no classifier-free guidance.
- References: up to 16 images per editing request.
- Architecture: a 33.1B-parameter DiT, the H3 design without the audio layers.
What it takes to run
The DiT weighs 66.28 GB in BF16, plus the MiniMax-H3 text encoder at another 66.7 GB, downloaded separately. An optional prompt expander (H3-Image-PE, NVFP4, 17.9 GB) helps because the model expects a detailed English description. Supported frameworks are SGLang, vLLM-Omni, diffusers and ComfyUI. The vLLM-Omni recipe lists NVIDIA H200, RTX PRO 6000 Blackwell and AMD MI350X among others.
It is also on the API. Per Neural Brew, a 1024x1024 image costs about $0.022 and a 4096x4096 image about $0.069 in standard mode, with 50% off images above 16.6 megapixels until November 17 (Neural Brew).
The license is the story
The weights are not under an open-source license. The MiniMax-H3-Image Community License excludes the European Union, the United Kingdom, South Korea and the United States from its territory (license). Using the weights or their outputs there is not authorized. Other terms:
- Commercial products must prominently display "MiniMax-H3-Image".
- Companies above $1M in annual revenue, or offering a hosted service, need prior written permission.
- MiniMax says it will discuss separate licenses for excluded territories.
Pruna's distilled variant followed the same pattern (Pruna P-H3-Image: fast image model, but not for the EU). The API and Hailuo AI fall under their own terms of service, which is a separate question.
Why it matters
Technically it is a strong package: native 16 MP, 6 steps and many references in one model. But for teams in Europe or the US, the downloadable weights are off limits, so the realistic options are the API or a model with a license that covers your region. Here, "open weights" does not mean anyone can run it anywhere.
Produced with the help of AI, reviewed by an editor. How we make our articles
Reviewed and edited by the Neural Daily desk. About


