MiniMax says it has released H3-Image with open weights on Hugging Face, claiming 4K images in 10 to 20 seconds, according to @MiniMax_AI. The company says the model is built on its new H3 line and is available now through its API, MCode and Hailuo AI.

What the API supports

MiniMax's image generation guide describes H3-Image as handling both text-to-image and image editing. Text-to-image takes a prompt alone. Editing takes between 1 and 16 reference images.

Output size can be set as a width by height value with a total area between 655,360 and 20,000,000 pixels, or chosen by aspect ratio plus a 1K, 2K or 4K tier, with 4K as the default. A square 4K image is 4096 by 4096 pixels, and 16:9 at 4K comes out at 5472 by 3072. Each request returns 1 to 10 images in jpeg, png or webp, and returned links stay valid for 24 hours.

Prompt enhancement is on by default for text-to-image and can be turned off. Editing always uses it. A MiniMax-H3-Image-Highspeed variant has the same parameters and limits, and developers switch to it by changing the model name in the request.

Three things are not supported at launch: preview images, mask-based local editing and transparent backgrounds.

For another open-weight image model to compare against, see our earlier report on Qwen-Image-2.1.

Pricing and the launch discount

MiniMax's pricing overview lists token prices in yuan per million tokens:

Per 1M tokens (yuan)

H3-Image

H3-Image-Highspeed

Input text

15

15

Input image

10

20

Output image

25

50

Per-image list prices for standard H3-Image are 0.1430 yuan at 1K, 0.1531 yuan at 2K and 0.4522 yuan at 4K. Highspeed 4K is listed at 0.8550 yuan per image. An editing request with three 1K reference images and 4K output costs 0.5211 yuan on the standard model.

The 50% discount applies to H3-Image and Highspeed requests whose final output is at least 16.6 million pixels, running until 2026-11-18 00:00 Beijing time. The post describes it as applying to 16 million pixel-plus generations, so the pricing page's 16.6 million threshold is the one that decides eligibility. At that threshold, standard 4K drops to 0.2261 yuan per image, Highspeed 4K to 0.4275 yuan and a 4K edit to 0.2605 yuan. A 1K or 2K image falls below the threshold and is not discounted.

Billing warning for developers

Both image endpoints are not idempotent. The docs say a timed-out request still generates and is billed, and a retry bills again. The docs advise turning off automatic SDK retries, which matters more for paid image jobs than the discount does.

Our read: the 4K price and the retry billing rule are the details to check before building on H3-Image, more than the headline speed claim.

Sources