MiniMax has announced H3-Image, an image model built on its H3 video model, and has released the model weights on Hugging Face. The model is available now through MiniMax’s API, its MCode agent and the Hailuo AI app.
One caveat first: the weights are public, but under MiniMax’s own community licence, not an open-source one. The repository was still locked when we first checked on the afternoon of 10 November and opened later that day. The licence does not cover the European Union, the United Kingdom, South Korea or the United States, where MiniMax asks would-be users to contact it for a licence. Elsewhere, companies with more than $1 million in annual revenue, and anyone offering the model as a hosted service, need MiniMax’s prior written permission, and the model and its outputs may not be used to improve other AI models.
What the API offers
MiniMax’s image generation guide documents two endpoints, plus a faster MiniMax-H3-Image-Highspeed variant with the same parameters:
- Text to image:
POST /v1/images/generations - Editing:
POST /v1/images/edits, with 1 to 16 reference images - Size: output area up to 20,000,000 pixels, aspect ratios from 1:3 to 3:1, or area tiers of 1K, 2K and 4K. 4K, about 4096 by 4096, is the default.
- Batch: 1 to 10 images per request, prompts up to 32,000 characters.
MiniMax says 4K images take 10 to 20 seconds and look free of the usual “AI look”. Those are its claims; we have not run it.
What it costs
Billing is per token. The pricing page lists the standard model at $2.30 per million text tokens in, $1.50 per million image tokens in and $3.80 per million image tokens out; Highspeed doubles the image rates. MiniMax’s worked examples put a standard 1K text-to-image at about $0.022 and a 4K one at about $0.069, or about $0.034 with the launch discount. That discount, 50% off outputs of at least 16.6 megapixels, runs until 17 November 2026 at 16:00 UTC.
One thing to set before you call it
The guide warns that neither endpoint is idempotent. If your client times out or disconnects, the server still finishes and bills the images, and a retry generates and bills them again. The OpenAI SDK retries twice by default, so MiniMax tells you to set max_retries=0 and a client timeout of at least 600 seconds.
This is a different model from Pruna’s P-H3-Image, a third-party image model built on the open H3 video weights with a territorial licence. For how H3 itself compares as a video model, see our MiniMax H3 vs HunyuanVideo episode, and for the closed and open image field, Seedream vs Qwen-Image.
Why this matters
At MiniMax’s listed rates, 4K output with multi-reference editing costs a few cents per image, which makes it easy to try against the image API you use now. Because the endpoints bill on disconnect, turn off automatic retries in your client before you wire it into a pipeline, or a flaky network will charge you twice. If you need self-hosting, wait for the weights and the licence text; H3’s video weights came under a community licence that excludes the US and EU, and the image release may carry similar terms.
The Campfire
CommentsThanks. Your comment is waiting for the desk. Once it is approved it appears here within about a minute.
Add a comment