MiniMax has released H3-Image, a still-image model for designers, agencies and marketing teams that want to generate pictures or edit from references. Its release notes list November 10, 2026, and its model documentation describes native 4096 × 4096 output, without a separate upscaling step.
The community licence sets regional and commercial limits that matter before a team commits to hardware or client work. Check eligibility first, then compare local and hosted options.
Check who can use it—and where the images will go
Downloadable weights are not unrestricted open-source permission. The H3-Image community licence sets three important boundaries:
Territory: the standard grant excludes the United States, United Kingdom, European Union and South Korea. MiniMax invites requests for a separate licence.
Business use: prior written authorization is required if you and/or your affiliates exceed $1 million in annualized revenue, or if you provide a hosted service to third parties, regardless of revenue.
Outputs: the territorial restriction also covers using, distributing and displaying generated images. Running the model in an eligible country does not, by itself, settle a worldwide campaign’s permissions.
The same agreement requires the model name on commercial interfaces and clear AI disclosure for public generated content. MiniMax claims no ownership of outputs, but that does not remove the use restrictions. These are adoption checks, not an exhaustive summary of the agreement.
Hosted access avoids downloads, not eligibility questions
MiniMax documents a hosted image-generation API, and Comfy confirms both native local support and API-based Partner Nodes. The latter lets ComfyUI users work through a remote service without downloading model weights.
The public documents do not clearly resolve how the community agreement interacts with every hosted service’s terms, regional access or multinational client arrangement. Do not assume that paying for an API establishes an exemption. Get confirmation of the intended use before committing client work.
For local use, Comfy’s current tutorial lists four component downloads: 21 GB, 25.28 GB, 7.34 GB and 1.17 GB. Storage calculation: these total 54.79 GB if none is already present, before software, caches and outputs. That is disk space, not a graphics-memory requirement or a guarantee that a particular GPU can run the workflow.
The Comfy text-to-image template starts at roughly four megapixels to reduce memory pressure. It uses compressed model components and a smaller prompt expander—a model that rewrites the instruction—than MiniMax’s official optional expander. Its text-to-image rewriting is off by default; the hosted API’s rewriting is on by default and mandatory for edits. Those configuration differences matter when comparing results; a shared model name is not a controlled comparison.
Check the editing task, not just the resolution
The API guide supports one to 16 reference images, but lists masked edits and transparent backgrounds as unsupported. If a job needs a precisely selected region changed or a transparent product cutout, retain another editing tool for that step.
Native resolution describes the output dimensions; it does not establish accurate lettering, faithful product details or reliable preservation of every reference. Review those details before approving an asset. This release concerns still images, distinct from the H3 video workflow covered in our July analysis.
Budget from usage, not a launch-price example
As of November 10, MiniMax’s API pricing is based on text and image tokens, its usage-accounting units, rather than a fixed fee per picture. These prices are in US dollars per million tokens:
Variant | Input text | Reference images | Generated images |
|---|---|---|---|
Standard | $2.30 | $1.50 | $3.80 |
Highspeed | $2.30 | $3.00 | $7.60 |
The listed promotion halves the total request cost for either variant when the final delivered image has at least 16,600,000 pixels. It ends on November 17, 2026 at 16:00 UTC; that endpoint is excluded.
Pixel calculation: 4096 × 4096 = 16,777,216 pixels, which qualifies. A 3840 × 2160 image has 8,294,400 pixels and does not. This assumes those are the delivered dimensions: a “4K” label alone is insufficient.
The published Standard example for 4096 × 4096 text-to-image is $0.06879 before the promotion and $0.03439 during it. Those are illustrative amounts, not fixed prices or averages. Actual token usage determines the bill.
There is another budgeting detail: the generation API reference says successful images remain billable if the client disconnects, and retrying creates another generation and charge. An integration should not blindly retry an ambiguous timeout. A technically successful image can also be creatively unusable, so track spending per accepted asset and the time needed for corrections—not just the cost of each generation.
Make the first evaluation small and specific
Once eligibility is clear, choose one real task—such as a campaign concept or a reference-based revision—and cap the evaluation budget. Record the access route, output size and rewriting settings, then assess the result against the actual brief. Move into routine work only when the permissions, editing effort and cost per usable asset fit the job.