MiniMax releases H3-Image weights, with substantial geographic and commercial limits
MiniMax has released H3-Image, an image-generation and editing model with downloadable weights. The company’s Chinese announcement and model license are dated November 10; its English announcement is dated November 11. MiniMax says the system can make a 4096-by-4096 image in under 10 seconds and work with as many as 16 reference images. The speed figure is a company claim, not a result independently reproduced for this article. MiniMax Chinese announcement English announcement
The release combines text-to-image generation and reference-image editing in one checkpoint. Its model card lists a 33.1-billion-parameter diffusion transformer and six sampling steps, with direct output of approximately 16 megapixels across supported aspect ratios. The listed image-model components total approximately 68.65 GB, while the separate H3 text encoder adds 66.7 GB. File size is not the same as the GPU memory required to run it, and no consumer-hardware performance claim follows from those numbers. Model card
The most consequential qualification is the license. The community agreement excludes the United States, European Union, United Kingdom and South Korea from its standard territorial grant. It invites interested users in excluded territories to seek a separate license. It also requires prior written authorization for third-party hosted services or when a licensee and its affiliates exceed $1 million in annualized revenue. Those restrictions make “open weights” a more accurate description than unrestricted open source. H3-Image license
MiniMax also offers hosted access through its API platform and consumer products. Availability of those services should not be confused with permission to self-host under the community agreement. Developers considering a deployment need to assess the applicable service terms or model license for their specific use. MiniMax Chinese announcement English announcement
The technical release expands the choices for high-resolution generation and editing. Its practical value will depend on reproducible speed, image quality, hardware costs and permission to use it in the intended market.
Sources and citations
Revision history
Published .