MiniMax released H3-Image on November 11, an image model with open weights. It handles text-to-image, generating from reference pictures, and editing existing images. A 4096×4096 image comes out in under 10 seconds, and you can reference up to 16 pictures at once.
What happened
H3-Image is built on MiniMax's H3 model, with generation and editing combined in one model. MiniMax highlights three things:
- Fast: A 4K image in under 10 seconds. MiniMax says it can reach up to five times the speed of comparable mainstream products.
- Looks real: Less of the overly smooth, plastic texture common in AI images, with more natural detail.
- Follows instructions: The strongest prompt adherence among open-weights models. It accepts up to 16 reference images, which is handy when you want separate references for the person, the outfit and the setting.
The model outputs roughly 16 megapixels directly, so you don't need an upscaler. Besides squares, it supports common ratios like 4:3, 16:9 and 21:9.
Speed comparison
The chart comes from MiniMax's blog and shows median generation time in seconds, where lower is faster. At 4K, H3-Image takes about 16 seconds in standard mode and about 10 seconds in fast mode, while Nano Banana 2.1 takes around 50 seconds. At 1K the gap is much smaller.
How to use it
| Way in | Who it's for | What you need |
|---|---|---|
| Hailuo AI web app, MiniMax Agent desktop app | Casual users who just want images | A MiniMax account |
| MiniMax API | Anyone who wants their AI agent to generate images in bulk | An API key, billed per image |
| Self-hosting the weights | Teams with data-center-class GPUs | Download from Hugging Face. Supports vLLM-Omni, SGLang, diffusers and ComfyUI |
On price, MiniMax says it's 50 to 70 percent of comparable mainstream services. During launch week, 4K and larger requests get a 50% discount.
To have your AI agent make an image for you, send it this:
Help me generate an image with MiniMax's H3-Image model.
1. First check whether I have a MiniMax API key on this computer (environment variable or config file). If not, stop and tell me where to apply at platform.minimax.io. Don't sign up or pay for anything on my behalf.
2. Call H3-Image following the official image generation docs. Size: [e.g. 5472x3072, 16:9 landscape]. Content: [describe the scene in one or two sentences]
3. The official guidance is to describe the whole image in detailed English. Expand my description into an English prompt and show it to me first. Generate only after I confirm.
4. Save the image to [folder path], then tell me the file name and roughly how much it cost this time.
Watch out
- Open weights doesn't mean free for commercial use. It uses MiniMax's own community license, not a common open-source license, so read the LICENSE before doing any commercial work.
- Local hosting is hardware-hungry. The main image model is about 66 GB, and you'll also need to download a text encoder of around 67 GB separately. MiniMax tested on pro cards like the H200 and RTX PRO 6000. A typical computer won't run it.
- Some settings are fixed. When self-hosting, the sampling step count must be 6, and negative prompts aren't supported. To keep something out of the picture, describe what you want in the positive prompt instead.
- The 10 seconds is MiniMax's number. Real speed depends on queue time and your network, and the "5x" claim is measured against specific competitor tiers.
Comments (0)
No comments yet. Start the conversation.