MiniMax has open-sourced MiniMax-H3-Image, a 33.1-billion-parameter image generation and editing model that can directly produce roughly 16-megapixel images in just six sampling steps — no separate super-resolution model needed. Why it matters: high-resolution output is usually where open image models fall behind closed services; a downloadable model that skips the upscaling step changes the cost and workflow math for teams willing to self-host.
November 10, 2026
What happened
According to Hugging Face listings reported on November 10, MiniMax open-sourced MiniMax-H3-Image under a community license agreement. A single set of weights supports both text-to-image generation and image editing, and the model accepts up to 16 reference images as input, according to PANews reporting on the release.
The model is also available through the MiniMax API and the Hailuo AI platform, so teams that do not want to run the weights themselves have a hosted route.
The details that matter
Reporting on the release describes a diffusion Transformer with about 33.1 billion parameters and roughly 68.65 GB of weights. It can generate images directly at about 16 megapixels, with resolutions ranging from 4096×4096 to 6272×2688 depending on aspect ratio. Sampling was distilled down to six steps, which is the efficiency claim to watch: fewer steps usually means faster and cheaper generation, if quality holds up in independent testing.
MiniMax says the model has the strongest instruction-following capability among open-source models. That is a vendor claim, not an independently verified ranking — as of November 10 there was no entry for MiniMax-H3-Image on the major public image leaderboards covered in early analysis of the release, so buyers should run their own prompt tests before treating the claim as settled.
The license catch
“Open-sourced” here means a community license, not an unrestricted one. Analysis of the MiniMax-H3-Image Community License Agreement notes that separate prior written authorization from MiniMax is required if annualized revenue exceeds one million US dollars, or if you provide hosted services to third parties. Commercial interfaces also carry branding requirements in the license analysis. Teams planning a paid product on top of the weights should read the full agreement before building on it.
Why it matters
For creators and small studios, the practical appeal is straightforward: one model for generation and editing, reference-image support, and native high resolution without bolting on an upscaler. For the wider market, H3-Image is another sign that the open-weight race is moving from “can it match closed models at all” to specific workflow wins — resolution, steps, and editing in a single checkpoint. Everything AI, tested means we will believe the instruction-following crown when independent benchmarks, not the model card, say so.
Frequently asked questions
What is MiniMax-H3-Image?
An open-weight image generation and editing model from MiniMax, released on November 10, 2026, with 33.1 billion parameters and support for up to 16 reference images.
How large an image can it generate?
About 16 megapixels directly, with resolutions reported from 4096×4096 to 6272×2688, without a separate super-resolution step.
Can I use it commercially for free?
It is released under a community license. Reporting on the agreement says separate written authorization is needed above one million US dollars in annualized revenue or for hosted services — read the license before commercial use.
Sources: PANews, Phemex News, OrcaRouter analysis, vLLM Recipes