MiniMax Releases Open-Weight H3-Image Model — 4K Images in Under 10 Seconds

Chinese AI lab MiniMax has open-sourced a 33-billion-parameter image model that generates 4K-resolution images in under 20 seconds — dropping it on Hugging Face on November 11, 2026 for anyone with enough storage and GPU memory to run it. The catch: the license excludes the United States, European Union, and United Kingdom.

MiniMax released H3-Image alongside a technical brief explaining the architecture and benchmarks.

What H3-Image Is

H3-Image is a diffusion transformer — a model architecture that iteratively refines random noise into a coherent image, guided by a text prompt or reference images. At 33.1 billion parameters, it sits at the large end of open-weight image models currently available.

Key specs:

• Resolution: Up to 4K×4K (16 megapixels) per image
• Speed: 10–20 seconds per image at full resolution
• Reference images: Accepts up to 16 reference images for consistent styling — useful for maintaining a character's appearance or a brand's visual identity across a series of images
• Model size: 135 GB download
• Compatible frameworks: SGLang, Hugging Face diffusers, and ComfyUI (the most widely-used open-source image generation interface)

The 135 GB size means this is not a model you run on a consumer laptop. You need a multi-GPU workstation or a cloud GPU instance with enough VRAM to hold the weights — roughly 2–3 A100-class GPUs depending on precision settings.

The License Restriction

MiniMax's release license explicitly excludes commercial and research use in the US, EU, and UK. This is an unusual restriction for an "open" model release, and it has immediate practical implications.

Developers and companies in those regions cannot use H3-Image for commercial products under the current license terms. Personal, non-commercial experimentation is a gray area that will likely be tested. The restriction appears to reflect MiniMax's regulatory and competitive positioning — the model can be used freely across much of Asia, Latin America, Africa, and the Middle East, while keeping it out of the jurisdictions most likely to pursue intellectual property claims or regulatory action against Chinese AI companies.

For context, MiniMax is the lab behind the Hailuo video model and the MiniMax-Text-01 language model. The company has been methodically releasing competitive open-weight models across modalities.

Why It Matters for Creators

The 16-reference-image input is the most commercially interesting capability here. Existing open-weight image models typically accept one or two reference images for style or subject consistency. Supporting up to 16 means:

• Character consistency across scenes — a filmmaker or game developer can feed reference shots of a character from multiple angles and get consistent output
• Brand style guides encoded directly — a marketing team can provide the model with a brand's existing visual assets and generate new creative that stays on-brand
• Multi-shot product photography — an e-commerce team can provide multiple product angles and get consistent scene compositions

This is the capability that video production houses and game studios have been waiting for in an open-weight model, and it's why the geographic license restriction is commercially significant — studios in the US and EU won't be able to use this without MiniMax changing the terms.

What to Watch

Whether MiniMax updates the license terms for US, EU, and UK users will determine how widely adopted this model becomes in professional creative workflows. Competing open-weight image models from Stability AI and Black Forest Labs (the FLUX model family) are fully permissive in those regions, which gives them an advantage for commercial use even if H3-Image outperforms them on resolution and consistency benchmarks. Watch for independent benchmark comparisons against FLUX.1 and similar models in the coming weeks.