MiniMax announced H3-Image’s publicly available image-generation and editing weights on November 11.

The big change

  • MiniMax says one checkpoint generates and edits images, accepts up to 16 references and produces roughly 16-megapixel results without upscaling.
  • Running the model requires planning accelerator memory. Hugging Face’s documented setup moves components between CPU and accelerator because they cannot fit together on its example GPU.
  • The community license limits both deployment and outputs geographically. Download access does not mean permission to operate a service everywhere.

From references to a new picture

MiniMax’s model card describes encoding instructions and reference images, predicting an image representation, then decoding it into pixels. The result is generated rather than a guaranteed unchanged collage.

Its example combines an empty room and nine furniture references into a furnished room. This illustrates the intended workflow, not independently verified reliability. Check individual objects and placement before using a result.

BIG CHANGE has not run the model or tested its performance.

The launch claims 4096-square generation in under 10 seconds without stating its hardware or full timing conditions. That claim cannot establish performance on a reader’s setup.

What running it involves

Hugging Face’s Diffusers documentation describes a transformer and text encoder each occupying about 62 GB in bfloat16, a numerical format used to store model values. Together they exceed an 80 GB GPU’s memory.

Its example automatically transfers each component onto the accelerator when needed. This avoids keeping both there simultaneously; it does not eliminate the need to hold and move them. The documented 4736-by-3552 run peaks at about 65 GB. That is an example configuration, not a universal minimum or a promise that a laptop can run the model.

For a deployment decision, compare the complete workflow and hardware configuration, rather than treating a component’s size as the whole system requirement.

Where the license applies

The license, dated November 10, excludes the EU, UK, Republic of Korea and US. It prohibits using or displaying the works or outputs outside its territory; people there can request separate licensing.

Prior written authorization is also required when the user and/or affiliates exceed $1 million annualized revenue, or for any third-party hosted service. The hosting condition applies independently of revenue. Redistribution and downstream safeguards have further requirements; this is not an unrestricted grant.

These are community-license conditions. Do not assume that separate hosted-product contracts have identical terms; check the intended service’s agreement and availability.

Paying for the hosted API

MiniMax’s pricing page lists Standard text input at $2.30, image input at $1.50 and image output at $3.80 per million tokens. Its 4096-square text-to-image example costs $0.06879 before the offer, or $0.03439 during it.

The 50% discount covers the total request cost when the final delivered image contains at least 16.6 million pixels, until November 17 at 16:00 UTC, exclusive. These are token-based examples, not fixed per-image charges. Changing inputs can change the bill, so compare representative requests rather than multiplying one example price across every task.