MiniMax released MiniMax H3-Image on November 11, 2026: an open-weight model for text-to-image, multi-reference generation and image editing, built on the same family as the MiniMax H3 video model. MiniMax says it renders 4K images in under 10 seconds and accepts up to 16 reference images. Here are the specs, the license, the hardware you need, and how it fits into an image-to-video workflow.
MiniMax H3-Image in one minute
- Released: November 11, 2026, by MiniMax.
- What it does: text-to-image, reference-based generation (up to 16 reference images) and image editing.
- Output: about 16 megapixels at every aspect ratio, from 4096 × 4096 (1:1) to 6272 × 2688 (21:9).
- Speed: 4K in under 10 seconds, according to MiniMax, with a Highspeed mode.
- Weights: on Hugging Face (
MiniMaxAI/MiniMax-H3-Image) and ModelScope, under the MiniMax-H3-Image Community License. - Hosted access: MiniMax API Platform, MiniMax Code and Hailuo AI, with 50% off 4K-and-above requests during launch week.
Architecture and specs
According to the model card, H3-Image is a 33.1-billion-parameter dense single-stream diffusion transformer (DiT) with 50 layers. It is distilled to 6 sampling steps and runs without classifier-free guidance, which is one reason it can be fast at very high resolution. It reuses the MiniMax H3 text encoder, based on Qwen3-VL, so the image model and the video model read prompts the same way.
| Spec | MiniMax H3-Image |
|---|---|
| Parameters | 33.1B, dense single-stream DiT |
| Sampling steps | 6 (distilled, no CFG) |
| Native output | About 16 MP (4096 × 4096 to 6272 × 2688) |
| Working range | About 4 to 20 MP |
| Reference images | Up to 16, referenced in the prompt as <Picture 1>, <Picture 2>… |
| Negative prompt | Not supported |
| Weights size | 68.65 GB (BF16) plus a separate text encoder of about 66.7 GB |
| Serving | vLLM-Omni, SGLang, diffusers, ComfyUI |
An optional prompt expander, MiniMax-H3-Image-PE, ships separately in NVFP4 (17.9 GB). It rewrites short prompts into detailed ones before generation.
Is MiniMax H3-Image open source?
It is open-weight, not open source in the strict sense. The weights are public and downloadable, but they come under the MiniMax-H3-Image Community License Agreement, and the model card says some uses need a separate written authorization requested through a form. Read the license before using it in a commercial product.
How fast and how good is it?
MiniMax claims up to 5 times the speed of mainstream alternatives, state-of-the-art prompt adherence among open-weight models, and pricing at 50 to 70% of comparable hosted offerings. The announcement shows no benchmark numbers, so treat these as company claims until independent tests are out. The hardware bar is real: with the text encoder, you are downloading well over 130 GB, which puts local use in data-center GPU territory rather than on a laptop.
Using H3-Image for image-to-video
The best AI videos usually start with a strong first frame. Because H3-Image shares the text encoder of MiniMax H3, it is a natural fit for a two-step workflow:
- Generate a sharp key frame with consistent characters and products, using several reference images.
- Animate it with an image-to-video model such as MiniMax H3 or Seedance 2.5.
In Nano Banana Video Studio you can already run that workflow end to end in the browser: create the frame with Nano Banana 2.1 or GPT Image 2.6, then animate it with MiniMax H3, with no GPU to rent.
MiniMax H3-Image vs Nano Banana 2.1 and FLUX 3
| MiniMax H3-Image | Nano Banana 2.1 | FLUX 3 Image | |
|---|---|---|---|
| Maker | MiniMax | Black Forest Labs | |
| Weights | Open (community license) | Closed | Open variants (Klein) |
| Max references | 16 | 14 plus a reference video | 10 |
| Resolution | About 16 MP | 1K, 2K, 4K | Up to 4K |
Reference counts and resolutions are the vendors' own figures.
FAQ
When was MiniMax H3-Image released?
On November 11, 2026.
Where can I download MiniMax H3-Image?
On Hugging Face as MiniMaxAI/MiniMax-H3-Image and on ModelScope. The text encoder is downloaded separately from MiniMaxAI/MiniMax-H3.
Can MiniMax H3-Image make videos?
No, it makes and edits still images. For video, pair it with the MiniMax H3 video model or another image-to-video model.
Is MiniMax H3-Image free for commercial use?
Not automatically. It uses the MiniMax-H3-Image Community License, which requires separate authorization for some uses. Check the license terms on Hugging Face.
Sources: MiniMax announcement, November 11, 2026; Hugging Face model card. Speed and quality figures are MiniMax's claims; status as of November 11, 2026.
