HunterAlphaHub
OpenRouter model reference Facts from the public catalogue, dated and labelled
Jev Laya Fable 5.2 Verified
2026-11-13

Model reference · verified 2026-11-13

MiniMax H3-Image, specified

MiniMax H3-Image is a 33.1B-parameter diffusion transformer built from the H3 stack. It combines text-to-image and reference-based editing in one checkpoint, generates roughly 16 megapixels without an upscaler, and runs in six sampling steps without classifier-free guidance. This page is the part a model card makes you assemble yourself: the file sizes, the API limits, the per-image bill and the licence boundaries.

Launch post 2026-11-11 Read 2026-11-13 Hugging Face model card →

The specification

FieldValue
Model
MiniMax-H3-Image
Highspeed variant: MiniMax-H3-Image-Highspeed. Same parameters and limits.
Released
2026-11-11
Weights, launch post and API documentation published together.
Architecture
Dense single-stream DiT
33.1B parameters, 50 layers, hidden size 5,376, 56 attention heads.
Text encoder
Qwen3-VL, unchanged
The full MiniMax-H3 text encoder; the DiT reads hidden states from layer 50.
Image VAE
CNN encoder + 24-layer ViT decoder
16× spatial downsampling and 24 latent channels.
Sampling
6 steps · no CFG
The local API rejects other step counts and guidance scales.
Weights
68.65 GB BF16
Plus a separate 66.7 GB H3 text encoder.
Optional prompt expander
17.9 GB NVFP4
MiniMax-H3-Image-PE for local text-to-image prompt expansion.
Licence
MiniMax-H3-Image Community License Agreement
Open weights, not an OSI open-source licence.
Source huggingface.co/MiniMaxAI/MiniMax-H3-Image; minimax.io/blog/minimax-h3-image Verified 2026-11-13 Architecture, model files and sampling limits: model cardRelease date, API and pricing: MiniMax launch post and platform docs

Where the parameters sit

H3-Image is H3-Base with the audio branch removed. The text prompt and any reference images are encoded by the H3 text encoder; reference images also pass through the image VAE. Those tokens are packed into one sequence for a single-stream transformer, which predicts the latent for the target image. The VAE then decodes that latent.

The important number for a local operator is not only the 33.1B DiT. The text encoder is the full Qwen3-VL encoder from H3 and is a separate 66.7 GB download. The repository's own transformer folder is 66.28 GB, plus 2.34 GB for the VAE and small tokenizer and processor files. A complete local build starts around 135 GB before runtime state, and the optional prompt expander adds 17.9 GB.

Hugging Face model card for MiniMax-H3-Image showing the 68.65 GB BF16 DiT and the optional 17.9 GB NVFP4 prompt expander
The model card separates the 68.65 GB image model from the optional 17.9 GB prompt expander. The 66.7 GB text encoder is a third download, pulled from the base MiniMax-H3 repository.

Output sizes and the 16 MP target

The model is trained to work across roughly 4–20 megapixels and is at its best near 16 MP. The API exposes either a fixed width and height or an aspect ratio plus a 1K, 2K or 4K area preset. The 4K preset does not mean a square image every time: at 16:9 it returns 5472×3072, and at 21:9 it returns 6272×2688.

Aspect ratioCommon 16 MP output
1:14096×4096
4:3 / 3:44736×3552 / 3552×4736
3:2 / 2:35024×3360 / 3360×5024
16:9 / 9:165472×3072 / 3072×5472
21:96272×2688

Model card and API output-size tables, read 2026-11-13 Custom sizes are never rounded; out-of-range values return HTTP 400

API limits that affect product design

FieldValue
Prompt length
32,000 characters
With enhancement disabled, only roughly the first 4,096 tokens take effect.
Images per request
1–10
Default is one.
Reference images
1–16
JPEG, PNG, WEBP, HEIC or HEIF.
Request body
64 MiB
Base64 inflates the original file by about one third.
Output area
655,360–20,000,000 px
The server never scales or rounds a custom width and height.
Aspect ratio
1:3 to 3:1
Applies to both the preset and custom-size routes.
Output formats
JPEG, PNG, WEBP
JPEG is the default.
Not supported
Masks, transparency, previews
The API rejects these editing options.

The prompt number is the one most likely to surprise a builder. The API accepts up to 32,000 characters, but with prompt enhancement disabled only about 4,096 tokens of the prompt take effect. The local model card says the same thing in its own words. If a prompt is long because it carries layout instructions, measure the revised prompt that the API returns rather than assuming every character reached the model.

The 64 MiB request body is another real ceiling. Base64-encoded references inflate the original file, so sixteen large images can exceed the limit before the reference count does. Public URLs and uploaded file IDs avoid that multiplication; the API supports both.

Pricing: token rates and actual task examples

MiniMax bills image generation by tokens rather than by a flat per-image price. The table below gives the three rates that matter. The example costs are more useful for planning: they show that 4K is not ten times the price of 1K, and that adding references costs less than moving from Standard to Highspeed.

ModelText input /MImage input /MImage output /M
MiniMax-H3-Image
Standard
$2.30 $1.50 $3.80
MiniMax-H3-Image-Highspeed
Highspeed
$2.30 $3.00 $7.60
Example taskOutputStandardHighspeed
Text to image 1K · 1024×1024 $0.02179 $0.03529
Text to image 2K · 2048×2048 $0.02333 $0.03622
Text to image 4K · 4096×4096 $0.03439 launch · $0.06879 regular $0.06501 launch · $0.13001 regular
Image edit · 1 reference 1K $0.02654 $0.04525
Image edit · 3 references 1K $0.03563 $0.06037
Image edit · 1 reference 4K $0.03656 launch · $0.07312 regular $0.06837 launch · $0.13673 regular
Image edit · 3 references 4K $0.03957 launch · $0.07914 regular $0.07703 launch · $0.15407 regular

MiniMax pricing overview, read 2026-11-13 Launch discount ends 17 November 2026, 16:00 UTC Official pricing →

MiniMax H3-Image pricing chart comparing Standard and Highspeed task costs for 1K, 2K and 4K text-to-image and image editing
MiniMax's own example-cost table. The highlighted 4K rows carry the launch discount; the struck-through numbers are the regular prices that return after 17 November.

The licence is the largest constraint

The phrase “open weights” is accurate; “open source” is not. The MiniMax-H3-Image Community License Agreement grants a royalty-free licence only in the Applicable Territory, which excludes the European Union, the United Kingdom, South Korea and the United States. Hosted services and organisations above $1 million in annualised revenue need separate written authorisation. A commercial product must display the model name in its user interface.

FieldValue
Licence
MiniMax-H3-Image Community License Agreement
The model card links the agreement as LICENSE; it is not Apache-2.0 or MIT.
Applicable territory
Worldwide minus four markets
The EU, UK, South Korea and the United States are excluded from the weights licence.
Hosted service
Separate authorisation
Required before providing H3-Image or a derivative through an API or hosted product.
Revenue threshold
$1M annualised
A licensee above the threshold also needs separate written authorisation.
Outputs
MiniMax claims no rights
The user is responsible for the output and its downstream use.
Commercial marking
Name required
A commercial product using the model must display “MiniMax-H3-Image” in its UI.

Licence and acceptable-use policy read 2026-11-13 Licence text →

The hosted MiniMax API is a separate offering. A reader in an excluded market should not infer from an API account that the downloadable weights are licensed for local use there. That distinction belongs before the hardware decision, not after it.

Send us a link

A project built with Jev, a post, a video, a correction, a tip. We open the link, check it says what you said it says, and write the entry ourselves.

Required a link, and an email to reply to. Optional everything else.

Add context — all optional

One or two sentences about what it does, in your words. We write the entry ourselves.

Cost, latency, a benchmark — anything you measured. We attribute these to you.

We store what you type and email it to ourselves. No IP address, no user agent, no referrer — the same rule as the mailing list.

What happens to what you send →