Released November 10, 2026 · Open weights

MiniMax H3-Image
4K images in under 10 seconds

An open-weight model for text-to-image, multi-reference generation and image editing. It renders about 16 megapixels natively, with no upscaler, and accepts up to 16 reference images.

Official API docs Weights on Hugging Face
16 MPNative output, e.g. 4096×4096
< 10 sPer 4K image (vendor claim)
16Reference images per edit
6 stepsDistilled sampling, no CFG

What it can do

One checkpoint handles both generation and editing.

Native 4K output

About 16 MP at every aspect ratio, from 4096×4096 (1:1) to 6272×2688 (21:9). The API accepts up to 20 MP.

Multi-reference editing

Combine 1 to 16 reference images, for example an empty room plus nine pieces of furniture composed into one scene.

Photorealistic detail

Built to avoid the overly smooth “AI look”, with fine textures and natural lighting.

Prompt enhancement

Short prompts in any language are expanded into a detailed English description before generation. An open prompt-expander model is also available.

Two API tiers

MiniMax-H3-Image for quality and MiniMax-H3-Image-Highspeed for lower latency, with the same parameters.

Self-hostable

A 33.1B-parameter diffusion transformer. Runs with SGLang, vLLM-Omni, diffusers and ComfyUI.

Pricing

Token-based billing. Example per-image costs from MiniMax's pricing page.

TaskStandardHighspeed
Text-to-image, 1K$0.022$0.035
Text-to-image, 2K$0.023$0.036
Text-to-image, 4K$0.069 $0.034$0.130 $0.065
Edit, 1 reference, 4K output$0.073 $0.037$0.137 $0.068

4K requests are 50% off until November 17, 2026, 16:00 UTC, only when the delivered image is at least 16,600,000 pixels. Prices change; check the official pricing page.

Quick start with the API

The endpoints are OpenAI-compatible, so the OpenAI SDK works with a different base URL.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.minimax.io/v1",
    api_key="YOUR_MINIMAX_API_KEY",
    timeout=600,
    max_retries=0,  # requests are not idempotent
)

result = client.images.generate(
    model="MiniMax-H3-Image",
    prompt="An orange cat on a sunny windowsill, film texture",
    size="2720x1536",
    response_format="url",
)
print(result.data[0].url)
  • Not supported: transparent backgrounds, masked edits, negative prompts.
  • Sizes: aspect ratio between 1:3 and 3:1, area from 655,360 to 20,000,000 pixels.
  • Batch: 1 to 10 images per request, with optional streaming.

License notes for the open weights

Summary of the MiniMax-H3-Image Community License. Read the full text before building on it.

  • Excluded territories: the United States, the European Union, the United Kingdom and South Korea.
  • Written authorization required if you offer it as a hosted service to third parties, or if annual revenue exceeds $1M.
  • Attribution: commercial products must display “MiniMax-H3-Image” in the user interface.
  • No training other models with its outputs.

FAQ

Is MiniMax H3-Image free?

The weights are free to download under the community license. The API is paid from the first image. The Hailuo AI app gives new users free starter credits.

What hardware do I need to run it locally?

The BF16 transformer is about 66 GB and the text encoder about 67 GB. MiniMax lists single-GPU setups such as NVIDIA H200 and RTX PRO 6000 Blackwell, and multi-GPU sequence parallelism for lower latency. Community quantizations are available.

How good is its text rendering?

Early independent tests report more spelling errors in rendered text than leading proprietary models. Check text-heavy designs carefully.

Where can I use it besides the MiniMax API?

At launch it was available through fal.ai, Runware, Vercel AI Gateway, WaveSpeedAI and natively in ComfyUI.

More coming soon

Prompt examples, side-by-side comparisons with other image models, and a hands-on playground are in progress.