Image models
MiniMax H3-Image: 16-megapixel images with open weights
MiniMax has released H3-Image, a 33.1-billion-parameter image generation and editing model built on its H3 video model. It makes images of about 16 megapixels, takes up to 16 reference images and ships with open weights under a restricted licence.
On 10 November, MiniMax released H3-Image, an image model from its H3 family that handles both text-to-image generation and image editing in one checkpoint. The weights are public on Hugging Face, and the model is also available through MiniMax’s API, the Hailuo web app and a desktop app.
What it can do
The model card says H3-Image generates images of about 16 megapixels directly, without an upscaler, from 4096 x 4096 (1:1) to 6272 x 2688 (21:9). The API accepts any custom size up to 20 million pixels, or a preset of 1K, 2K or 4K, with 4K as the default.
Editing takes one to sixteen reference images, which is the feature most useful to anyone keeping a character or a location consistent across scenes. The API returns one to ten images per request in JPEG, PNG or WebP. MiniMax claims the strongest instruction following among open-source models, but the card gives no benchmark number for that.
The numbers
- Size: a diffusion transformer with 33.1 billion parameters; the main checkpoint is 68.65 GB in BF16.
- Speed: distilled to 6 sampling steps with no classifier-free guidance, which is how MiniMax keeps 16-megapixel output practical.
- Prompt expander: an optional 17.9 GB model, H3-Image-PE, rewrites short prompts. The API rewrites prompts by default, and edits always use the rewriting step.
- Serving: the card shows a one-GPU server and a four-GPU split for lower latency, and lists supported cards including the NVIDIA H200, RTX PRO 6000 Blackwell and AMD Instinct MI350X. It gives no minimum memory figure.
Limits and licence
The API guide lists three things as unsupported: preview images, masked edits and transparent backgrounds. The guide also warns that requests are not idempotent, so automatic client retries can generate, and bill for, the same job twice.
The licence is MiniMax’s own, not a standard open-source one. It excludes the European Union, the United Kingdom, the Republic of Korea and the United States, where you would need a separate licence from MiniMax. Written authorisation is also required if you or your affiliates have annualised revenue above 1 million US dollars, or if you offer the model as a hosted service to others. Read it in full before any commercial use. MiniMax’s speed and quality statements are its own and we have not tested them.
Our take
For people building interactive worlds, a 16-megapixel model with multi-reference editing is a practical tool for concept art, key art and consistent character sheets, and open weights mean a studio can run it on its own hardware. The catch is the licence: for teams in the UK or EU, the model is not freely usable. The bigger pattern is video models becoming the base for other creative tools, which we cover in generative interactive film.