MiniMax H3-Image: 4K Images, Open Weights and API Access
MiniMax H3-Image generates 4096×4096 images, edits with up to 16 reference images and offers downloadable weights. Here is what the official pages say about access, pricing and licensing.
On this page
- What is MiniMax H3-Image?
- How large is a native 4K image?
- How does editing with 16 reference images work?
- Standard or Highspeed: which model ID?
- What does the H3-Image API cost?
- Can you run the open weights locally?
- What do the official samples show?
- Where can you use H3-Image today?
- Frequently asked questions
- When did MiniMax H3-Image come out?
- Is MiniMax H3-Image open source?
- Does H3-Image support masked edits or transparent PNGs?
- How much does one 4K image cost?
- Is H3-Image the same as MiniMax H3?
- Sources checked
MiniMax H3-Image is available through MiniMax's API, MiniMax Code and Hailuo AI, with weights on Hugging Face. It generates 4096×4096 images and edits with up to 16 reference images. MiniMax's pricing page lists a 4K text-to-image example at $0.03439 at the launch rate.
- Model: one 33.1B-parameter checkpoint for text-to-image and reference editing, per the model card.
- Sizes: 4096×4096 at 1:1 to 6272×2688 at 21:9, plus custom sizes up to 20 MP.
- Access: MiniMax API, MiniMax Code and Hailuo AI; weights on Hugging Face under a custom community license.
What is MiniMax H3-Image?
H3-Image is the image member of MiniMax's H3 family. One checkpoint handles text-to-image and editing of one or more reference images, so a pipeline can keep one model instead of pairing a generator with a separate editor.
MiniMax's announcement pitches about 16 MP output, more natural detail and stronger instruction following than other open-weight models. Those are vendor claims; the API parameters, published sizes and prices are what you can plan around.
H3-Image is separate from MiniMax H3, the video model. For that API, the MiniMax H3 prompt guide covers its three-field structure and example clips.
How large is a native 4K image?
The 4K preset targets a 4096×4096 area (16.8 MP); at 1:1 that is the full image. MiniMax says it needs no upscaler. The model card lists about 16 MP sizes for each ratio:
| Ratio | 4K size |
|---|---|
| 1:1 | 4096×4096 |
| 4:3 | 4736×3552 |
| 16:9 | 5472×3072 |
| 21:9 | 6272×2688 |
The API also takes an exact WIDTHxHEIGHT. The image guide allows an area from 655,360 to 20,000,000 pixels and an aspect ratio from 1:3 to 3:1, returned without scaling or rounding. Use aspect_ratio plus area (1K, 2K, 4K) or size, never both; sending both returns 400.
How does editing with 16 reference images work?
Editing uses POST /v1/images/edits with a prompt and 1 to 16 reference images. Upload order sets the labels: the first file is <Picture 1>, the second <Picture 2>, and the prompt refers to them by name. The image guide documents the request format.
References can be JPG, JPEG, PNG, WEBP, HEIC or HEIF, up to 64 MiB each, 128 to 8192 px per side and no more than a 4:1 ratio. The request body is capped at 64 MiB; Base64 grows to about 4/3 of the file size, so several large references are easier to send as public URLs or mm_file:// IDs.
Prompt enhancement always runs on edits; data[].revised_prompt returns the prompt actually used.
Standard or Highspeed: which model ID?
The IDs are MiniMax-H3-Image and MiniMax-H3-Image-Highspeed. They use the same parameters and limits; their official price examples differ. MiniMax's announcement says 4K generation takes under 10 seconds, without naming a baseline, GPU or batch size. We did not run a speed test. Compare both IDs on your prompts before choosing.
What does the H3-Image API cost?
MiniMax bills by tokens; these are example request totals, not fixed per-image prices. The pricing page lists:
| 4K example | Standard | Highspeed |
|---|---|---|
| Launch | $0.03439 | $0.06501 |
| Regular | $0.06879 | $0.13001 |
The launch rate is 50% off before November 17, 2026 at 16:00 UTC when the delivered image has at least 16,600,000 pixels. The discount ends at that time. That covers 4096×4096 and the sizes above. At the launch example, 100 standard 4K images cost about $3.44; the regular example doubles that to $6.88.
Standard's token rates are $2.30 per million input text tokens, $1.50 per million input image tokens and $3.80 per million output image tokens. Highspeed raises input images to $3.00 and output images to $7.60. H3-Image-Context-IR prompt enhancement is $2.30 per million input tokens.
Can you run the open weights locally?
The weights are downloadable, but the license is a custom community license, not an open-source license. The agreement excludes the EU, UK, South Korea and the US from its Applicable Territory and requires prior written authorization when annualized revenue exceeds $1 million or when you provide a hosted service to third parties.
The H3-Image download is 68.65 GB, plus a separate 66.7 GB Qwen3-VL text encoder; a full install is roughly 135 GB. That is disk size, not VRAM. MiniMax's model card lists recipes for specific GPUs, including H200, RTX PRO 6000 Blackwell and AMD Instinct MI350X, and publishes no minimum VRAM. Check the card for a supported setup.
What do the official samples show?
These are MiniMax's model-card assets, cropped for layout. The model card has the originals.
A person hangs white laundry on a rooftop in warm evening light. A large moon sits below a billowing sheet, with rooftops and a reflecting puddle. Inspect the translucent cloth and warm light; this is a sample, not a photograph.
A woman in a red dress cuts a moon-shaped sphere on a plate at a white-clothed outdoor table under a starry sky. Ordinary tableware beside the impossible object and the warm/cool light define the scene.
The edit pair starts with a gray, clay-like lunar base: habitat, rover and astronaut. The result adds a black sky, strong shadows and metal-like materials. The arrangement stays recognizable; some fine shapes differ.
Where can you use H3-Image today?
MiniMax lists three hosted entry points: the API Platform, MiniMax Code and Hailuo AI. The API runs on the global and Chinese platforms; the weights are on Hugging Face and ModelScope. The announcement and model card list these.
The OmniaKey catalogue we checked on November 10 lists image models such as GPT Image 2.5 Sunburst, not H3-Image. The GPT Image 2.5 API guide shows how to call an image model through one key. You can compare the Seedream 5.5 launch notes or try models in the AI image generator.
For a hosted image API inside OmniaKey today, GPT Image 2.5 Sunburst is one current option. For H3-Image itself, use MiniMax's API and its launch discount through November 17.
Frequently asked questions
When did MiniMax H3-Image come out?
The announcement is dated November 11, 2026, and the license November 10, 2026. Both pages were live when we checked on November 10 in New York.
Is MiniMax H3-Image open source?
No. The weights are downloadable under a custom community license, not an open-source license; the agreement sets territorial and authorization limits.
Does H3-Image support masked edits or transparent PNGs?
No. The image guide lists preview images, masked edits and transparent backgrounds as unsupported. Editing takes 1 to 16 reference images plus a prompt and returns a full rectangular JPEG, PNG or WebP.
How much does one 4K image cost?
The 4K text-to-image example is $0.03439 on Standard and $0.06501 on Highspeed at the launch rate; regular examples are $0.06879 and $0.13001. Charges follow token usage.
Is H3-Image the same as MiniMax H3?
No. H3 is MiniMax's video model; H3-Image is a separate image checkpoint built on H3. OmniaKey's minimax-h3 entry is video, not this image API.
Sources checked
- MiniMax-H3-Image Community License Agreement, license date November 10, 2026.
- MiniMax: MiniMax H3-Image announcement, byline November 11, 2026.
- MiniMax: image generation pricing, checked November 10, 2026.
- MiniMax-H3-Image model card, checked November 10, 2026.
- MiniMax: image generation guide, checked November 10, 2026.
Related articles
MiniMax H3 Prompt Guide: Structure, Examples and Real Clips
Seven prompts, seven real clips: which cuts landed on time, what the sound fields did and which signs came out garbled.
Seedream 5.5: Release Date, New Features and API Access
ByteDance's newest image model is out, with steadier edit chains and real alpha output. Here is what shipped, where it runs, and what the release pages still leave open.
GPT Image 2.5 API: Generate and Edit Images in Python
One POST creates the task and one GET reads it back. Here is the request code, the sizes we ran on October 29, 2026, and one edit shown before and after.