MiniMax H3-Image: 16-megapixel open weights, but not for US or EU users

MiniMax has published H3-Image, a 33.1-billion-parameter open-weight model that handles both text-to-image generation and editing at about 16 megapixels, in six sampling steps. Its license, however, excludes the US, EU, UK and South Korea. The release was first flagged by the Stable Diffusion community; the details below come from the model card on Hugging Face.

A 33B diffusion transformer built for very large images

According to the model card, H3-Image is a diffusion transformer with 50 layers, a hidden size of 5376 and 56 attention heads. It reuses the unchanged Qwen3-VL text encoder from MiniMax-H3 and an image VAE with 16x spatial downsampling and 24 latent channels. The model is distilled to six sampling steps and needs no classifier-free guidance, which is how it keeps generation short for its size.

One checkpoint covers both text-to-image and editing, and it accepts up to 16 reference images. Output runs across aspect ratios from 4096×4096 (1:1) to 6272×2688 (21:9). MiniMax gives an effective range of roughly 4 to 20 megapixels, with the best results near 16. The BF16 checkpoint is 68.65 GB, so running it locally implies a high-memory GPU or quantization. The card does not list hardware requirements.

An optional prompt expander, MiniMax-H3-Image-PE, is 17.9 GB in NVFP4. We covered it separately in MiniMax open-sources a 27B prompt rewriter for its H3 image model.

A license that shuts out the largest markets

The weights ship under the MiniMax-H3-Image Community License Agreement, not a standard open-source license. The license text covers only an “Applicable Territory” defined as the world excluding the EU, UK, South Korea and the US. Use, distribution or display outside that territory is not authorized.

Two further triggers require prior written authorization from MiniMax, requested through its platform: annualized revenue above 1 million USD for the user and its affiliates, or offering the model to third parties as a hosted service. Commercial products must display the “MiniMax-H3-Image” name in the interface. Outputs may not be used to improve other AI models, other than MiniMax’s own. Hong Kong law governs the agreement.

Who can actually use it, and what is still unproven

For hobbyists and small studios outside the excluded regions, the practical change is a single model that edits and generates at print-scale resolution without a separate upscaling stage. Larger companies and anyone building a service on top of it must apply for permission first. Developers in the US or EU are outside the licensed territory altogether, which sets this release apart from fully permissive open-weight models. We saw a similar pattern in Viggle Remake and Kimi K3.1 Pro, where open weights came with license conditions.

Hosted access is the alternative. Third-party provider WaveSpeedAI lists H3 still-image endpoints at $0.02 per 1K image for text-to-image and $0.03 for editing, rising to $0.06 and $0.09 at 2K. Those are the provider’s prices for its own service, and its published tiers top out at 2K, well below the 16 megapixels in the weights. The article here does not establish that those endpoints are the same model as H3-Image.

The card also leaves gaps. It claims “the strongest instruction following among open-source models,” but publishes no benchmark scores to support this, and it states no release date. The claim should be treated as MiniMax’s own until independent comparisons appear. The card does not say how the quality of the six-step distilled model compares with undistilled alternatives.