MiniMax-H3-Image: Specs, License & How to Run

MiniMax-H3-Image is MiniMax’s official still-image checkpoint for text-to-image and image editing. Official hosts we opened on 2026-11-11: the ungated card MiniMaxAI/MiniMax-H3-Image (created 2026-11-03T06:34:36Z; last modified 2026-11-11T00:30:36Z; 400 likes / 667 downloads), LICENSE (MiniMax-H3-Image Community License Agreement, 10 November 2026), optional prompt expander MiniMaxAI/MiniMax-H3-Image-PE (30 likes / 0 downloads; created 2026-11-08T15:35:23Z), text encoder host MiniMaxAI/MiniMax-H3, the research blog MiniMax H3-Image: Bringing 4K Image Generation into the 10-Second Era (dated 2026-11-11), GitHub mirror MiniMax-AI/MiniMax-H3-Image, the vLLM-Omni recipe MiniMaxAI/MiniMax-H3-Image, and ComfyUI’s 10 November 2026 note MiniMax H3 Image is now available in ComfyUI. API gated was false. Hugging Face pipeline_tag is text-to-image. library_name is minimax-h3. Tags include image-to-image and image-editing.

This identification page hangs off open Midjourney-class alternatives. That mapping is editorial. We have no sourced bake-off against Midjourney.

It is not P-H3-Image (Pruna’s earlier still-image pack of MiniMax-H3, 8 steps, ~8 MP). Head-to-head facts: MiniMax-H3-Image vs P-H3-Image. It is also not Pika H3 Flex (video+audio adapters on MiniMax-H3). P-H3-Image already has P-H3-Image vs FLUX 3 Image Klein; this page does not duplicate that pairing.

The card’s “strongest instruction following among open-source models” sentence and the blog’s “under 10 seconds” / “5× the speed of mainstream alternatives” lines stay vendor.

What we verified (2026-11-11)

Field	Value	Source
Host	huggingface.co/MiniMaxAI/MiniMax-H3-Image (not gated)	HF API
Created / last modified	2026-11-03T06:34:36Z / 2026-11-11T00:30:36Z	same
Likes / downloads	400 / 667	same
License tag + file	other / minimax-h3-image-community-license-agreement	HF cardData + LICENSE
LICENSE date	10 November 2026	LICENSE
Pipeline tag	text-to-image	HF API
Base model tag	MiniMaxAI/MiniMax-H3 (finetune)	HF tags
Transformer class	MiniMaxH3DiTModel (50 layers, 56 heads, hidden 5376, audio_enabled false)	transformer/config.json
Params (card)	33.1B dense DiT	README
Model Zoo size (this repo)	68.65 GB	README table
Text encoder	MiniMax-H3 Qwen3-VL, ~66.7 GB, loaded from MiniMaxAI/MiniMax-H3 text_encoder/	README + model_index.json
usedStorage (this card)	204,002,730,228 bytes (~190 GiB)	HF API
Sampling	6 steps (6 NFEs), guidance_scale 1, no negative prompt	README
Resolution	Native ~16 MP without an upscaler; working range ~4–20 MP	README
Editing	Up to 16 reference images	README
Optional PE	MiniMaxAI/MiniMax-H3-Image-PE, NVFP4, 17.9 GB, 30 likes	Model Zoo + HF API

model_index.json still points the text encoder at MiniMaxAI/MiniMax-H3 text_encoder/. Hugging Face siblings on this card also list text_encoder/ shards plus transformer_diffusers/ and vae_diffusers/. That is why usedStorage (~190 GiB) is larger than the Model Zoo 68.65 GB label. We keep both quotes.

LICENSE date is 10 November 2026. The research blog HTML we opened is dated 2026-11-11. Hugging Face createdAt is 2026-11-03. Identification uses those three dates as they appear.

Native sizes (copied from the card)

The card says the effective working range is about 4 MP to 20 MP, and that it generally gives the best results at around 16 MP. Listed ~16 MP sizes (width × height):

Aspect ratio	Size
1:1	4096×4096
4:3 / 3:4	4736×3552 / 3552×4736
3:2 / 2:3	5024×3360 / 3360×5024
16:9 / 9:16	5472×3072 / 3072×5472
21:9	6272×2688

No upscaler is in that loop. Other aspect ratios at ~16 MP are allowed; those five rows are the card’s common list.

How to run (from official cards, not tested)

Card download:

hf download MiniMaxAI/MiniMax-H3-Image
hf download MiniMaxAI/MiniMax-H3 --include "text_encoder/*"
# optional prompt expander
hf download MiniMaxAI/MiniMax-H3-Image-PE

Recommended serving frameworks copied from the README: SGLang, vLLM-Omni, diffusers, ComfyUI.

Card SGLang one-GPU line:

sglang serve --model-path MiniMaxAI/MiniMax-H3-Image --host 0.0.0.0 --port 30000 --warmup-mode off

Card vLLM-Omni one-GPU line:

vllm serve MiniMaxAI/MiniMax-H3-Image --omni --host 0.0.0.0 --port 8000

Request contract copied from the README: /v1/images/generations (JSON) for text-to-image; /v1/images/edits (multipart) for references. size is required. num_inference_steps must be 6. guidance_scale must be 1. Other values return HTTP 400. negative_prompt is rejected. Prompts are truncated at 4096 tokens.

The MiniMax API turns a short prompt into a detailed English description with H3-Image-Context-IR before generation. Local alternative: optional MiniMax-H3-Image-PE. forkalt has no PE page; PE is an optional expander on this checkpoint, not a separate product listing.

ComfyUI: the 10 November 2026 Comfy blog says MiniMax H3 Image runs natively (T2I and Edit templates). Comfy-Org/MiniMax-H3-Image is a repack of these weights, not a third identification page.

Hosted platform.minimax.io and Hailuo are API / product SKUs. Blog Standard / Highspeed multipliers and “50–70% of comparable mainstream offerings” are not a statement about the weight files. forkalt has not served the shards.

Vendor speed and memory (self-reported)

Not a forkalt measurement. Copied labels only.

11 November 2026 MiniMax blog: a 4K (4096×4096) image “in under 10 seconds,” “at up to 5× the speed of mainstream alternatives.” Those comparators are unnamed.

vLLM-Omni recipe MiniMaxAI/MiniMax-H3-Image, H200 rows as printed there: 1 GPU with text-encoder offload 73.4 GiB peak / 3.2 s at 2048²; 1 GPU DLO (DiT + text encoder) 14.7 GiB / 8.3 s; 4 GPU Ulysses 81.3 GiB / 0.88 s. The same recipe’s intro also prints a 2720×1536 image “in about 0.8 s” on 4×H200. That recipe’s overview says “up to nine reference images” and “2K-class canvases.” Identification of the weights follows the MiniMax card (16 refs, native ~16 MP). The recipe rows stay vendor serving notes.

License

LICENSE is the MiniMax-H3-Image Community License Agreement (10 November 2026). Hugging Face cardData.license is other. We do not invent a more specific SPDX id. This is a different file from the August 2026 MiniMax H3 Community License on P-H3-Image.

Paraphrase only, not legal advice:

• Applicable Territory is worldwide excluding the European Union, the United Kingdom, the Republic of Korea, and the United States of America. Use outside that territory is not authorized by this agreement (contact MiniMax for a separate license).
• LICENSE §V.4: you may not use, reproduce, modify, distribute, or display the Works or their Outputs outside the Applicable Territory.
• §IV.1: Annualized Revenue of you and/or affiliates over USD 1 million, or providing Hosted Service(s) to any Third Parties, requires prior written MiniMax authorization (platform.minimax.io/console/license).
• §IV.2: prominently display “MiniMax-H3-Image” on commercial product UIs. §III.3 encourages “Powered by MiniMax-H3-Image”.
• You may not use the Works or Outputs to improve any other AI model other than MiniMax-H3-Image or its Model Derivatives.
• MiniMax-H3-Image-PE is part of MiniMax-H3-Image and follows this agreement. The LICENSE additional note says PE is derived from Qwen3.8-27B (Apache-2.0). That Apache grant on the starting weights does not replace this Community License for PE.

This $1 million / Hosted Service trigger is stricter than the $20 million yearly-revenue clause on the older MiniMax H3 Community License used by P-H3-Image. Do not mix the two files.

What we did not invent

• Photorealism / instruction-following / “under 10 seconds” / “5×” stay vendor.
• vLLM-Omni latency and peak-VRAM rows stay vendor. We have not measured 4096².
• We do not rank MiniMax-H3-Image against P-H3-Image, FLUX 3 Image Klein, Qwen-Image-2.1-Turbo, MOSS-Image, Iris-3B, or Midjourney.
• We do not treat P-H3-Image as a quantization of this card. Pruna published P-H3-Image from MiniMax-H3 before this official image checkpoint’s LICENSE date.

Related on this site: MiniMax-H3-Image vs P-H3-Image — P-H3-Image — P-H3-Image vs FLUX 3 Image Klein — FLUX 3 Image Klein — Qwen-Image-2.1-Turbo — MOSS-Image — Iris-3B — Fibo 2.0 (CC BY-NC) — Pika H3 Flex (same MiniMax-H3 family, video) — MiniMax M3.1 Flash (same lab, chat MoE, not an image DiT) — ComfyUI. Hub: Midjourney-class alternatives. Catalog: home. Corrections: contact.