What Is MiniMax H3-Image? MiniMax's Open-Weights 4K Image Model Explained (2026)
What Is MiniMax H3-Image? MiniMax's Open-Weights 4K Image Model Explained (2026)
Short answer: MiniMax H3-Image is MiniMax's new open-weights image model, announced on the official MiniMax blog dated November 11, 2026. It is built on the H3 family and unifies text-to-image, reference-based generation, and image editing in one checkpoint. MiniMax says it can produce a native 4K (4096 × 4096, about 16 MP) still in under 10 seconds, follow long prompts tightly, and work with up to 16 reference images. Weights are on Hugging Face at MiniMaxAI/MiniMax-H3-Image. Official hosted access is also on the MiniMax API, MiniMax Code, and Hailuo AI. If you want a browser image tool without standing up a large local stack, see HackAIGC's uncensored image generator.
This page is the model explainer. For the separate “can it do NSFW / what about filters?” question, see Can MiniMax H3-Image Do NSFW?. Do not confuse H3-Image with MiniMax H3 the video model — that is a different product, covered in How to use MiniMax H3 video and MiniMax H3 vs uncensored alternatives.
TL;DR
| Fact | Official detail (Nov 10–11, 2026) |
|---|---|
| What it is | Open-weights image generation and editing model in the MiniMax H3 family |
| Announcement | MiniMax blog dated November 11, 2026; Community License date November 10, 2026 |
| Core modes | Text-to-image, reference-based generation, and image editing in one checkpoint |
| Native resolution | About 16 MP without an upscaler; 1:1 is 4096 × 4096; other 16 MP sizes go up to 6272 × 2688 (21:9) |
| Speed (MiniMax's claim) | A full-resolution 4K image in under 10 seconds, “up to 5×” mainstream alternatives; API Standard “1.5–2×,” Highspeed “3–4×” |
| References | Up to 16 reference images |
| Sampling (model card) | Distilled to 6 sampling steps (6 NFEs), no classifier-free guidance |
| DiT size (model card) | 33.1B parameters; BF16 checkpoint listed at 68.65 GB; the MiniMax-H3 text encoder is a separate 66.7 GB download |
| Official hosts | Hugging Face weights, MiniMax API, MiniMax Code, Hailuo AI |
| Day-one workflows | Native ComfyUI support (Nov 10, 2026); fal.ai endpoints for text-to-image and edit (standard and Highspeed) |
| License | MiniMax-H3-Image Community License; the Applicable Territory excludes the EU, UK, South Korea, and the United States unless you get a separate license |
| NSFW? | Not answered here — see the NSFW / filter page |
What MiniMax H3-Image is
MiniMax describes H3-Image as an open-weight image model built on H3. The Hugging Face model card is more specific: generation and editing are unified in one checkpoint. A text prompt alone is text-to-image. Connecting one or more reference images turns the same model into an editor.
That is the important product split:
| Name | Modality | Typical job |
|---|---|---|
| MiniMax H3 | Video (the earlier open H3 family model) | Text/image-to-video and related video workflows |
| MiniMax H3-Image | Still images | Text-to-image, multi-reference composition, instruction-based edits |
They share a family and, according to the image model card, the same MiniMax-H3 text encoder (the full Qwen3-VL text encoder, unchanged). They are not the same weights and not the same output type.
What MiniMax claims it does well
The launch post organizes the pitch around three strengths. These are MiniMax's published claims, not independent lab scores from this site.
Speed and native 4K
MiniMax says 4K (16 MP, 4096 × 4096) stills usually cost much more compute than the 1K or 2K outputs most tools default to, and that H3-Image makes that resolution practical: a full-resolution 4K image in under 10 seconds, at up to 5× the speed of “mainstream alternatives.” On the API, MiniMax separately describes Standard mode as 1.5–2× as fast as those alternatives and Highspeed as 3–4×. The API model name for the faster hosted variant is MiniMax-H3-Image-Highspeed.
We are not republishing a stopwatch test of our own. Treat the “under 10 seconds / 5×” figures as vendor claims, and measure them on the hardware or API plan you actually use.
Image quality
MiniMax says H3-Image aims to reduce the overly smooth “AI look,” with richer textures and more realistic detail. That is a qualitative claim in the launch post. The public showcase on the model card includes photoreal photography, editorial and concept work, characters, illustration styles, and examples of text rendering (including logic-puzzle boards in Chinese and English). We are not inventing benchmark numbers on top of those demos.
Prompt adherence and references
MiniMax calls prompt adherence state-of-the-art among open-weight models and highlights support for up to 16 reference images. Official edit examples on the card include turning a raw 3D render into a film still, turning a game map into a first-person view, and composing ten furniture references into one room.
The model card also says the checkpoint expects a detailed English description of the whole image. The hosted MiniMax API expands shorter prompts (in any language, MiniMax says) with H3-Image-Context-IR before generation. A local optional companion, MiniMax-H3-Image-PE (17.9 GB, NVFP4), does similar prompt expansion if you self-host.
Model-card specs (no invented numbers)
Figures below are from the public MiniMax-H3-Image model card and the image generation guide.
Architecture and files
| Item | Published figure |
|---|---|
| Diffusion transformer | 33.1B parameters; 50 layers; hidden size 5376; 56 attention heads; FFN size 14336; 3D multimodal RoPE |
| Precision | DiT weights BF16; embedding, timestep, and output layers FP32 |
| Image VAE | CNN encoder + 2D ViT decoder (24 layers); 16× spatial downsampling; 24 latent channels; each transformer token covers 32 × 32 pixels |
| Text encoder | Full Qwen3-VL text encoder of MiniMax-H3, not shipped inside the H3-Image repo; loaded from MiniMaxAI/MiniMax-H3 (text_encoder/, 66.7 GB) |
| Sampling | Distilled to 6 steps; guidance_scale fixed at 1; negative prompts are not supported on the documented local/server interface |
| Listed checkpoint size | MiniMax-H3-Image 68.65 GB; transformer folder 66.28 GB; VAE 2.34 GB |
Local inference frameworks named on the card include SGLang, vLLM-Omni, diffusers, and ComfyUI. The card does not publish a single official “minimum VRAM” number for every consumer GPU, so this page does not invent one. Plan around the published file sizes plus the separate text encoder.
Output sizes
The card says the effective working range is about 4 MP to 20 MP, with the best results generally around 16 MP, generated directly (not upscaled). Common 16 MP sizes:
| Aspect ratio | Size (width × height) |
|---|---|
| 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 |
The official API guide adds that hosted output area can go up to 20,000,000 pixels, with area presets of 1K / 2K / 4K (targets 1024², 2048², and 4096²) or an exact WIDTHxHEIGHT. Prompt length on the API is listed as ≤ 32,000 characters; with prompt enhancement disabled, only about the first 4096 tokens take effect. Image edits accept 1 to 16 references. Preview images, masked edits, and transparent backgrounds are documented as not supported on the official API.
Where you can use it
| Surface | What it is | Notes |
|---|---|---|
| Hugging Face: MiniMaxAI/MiniMax-H3-Image | Open weights + model card | You also need the MiniMax-H3 text encoder; PE is optional |
| MiniMax API | Official hosted text-to-image and edits | Global platform.minimax.io / api.minimax.io; CN counterparts on .cn. Standard and Highspeed model names |
| Hailuo AI and MiniMax agent / desktop | Official apps named on the model card | Web: hailuoai.video (global) / hailuoai.com (CN); desktop: agent.minimax.io / agent.minimax.cn |
| MiniMax Code | Official product listed on the launch post | Check the live product for current access |
| ComfyUI | Native open-weights graphs, plus Partner Nodes | Official Comfy blog, November 10, 2026: 6-step sampling, same 16 MP size table, templates for T2I and edit |
| fal.ai | Hosted text-to-image and edit, plus Highspeed variants | fal's launch page (dated November 10, 2026) also describes a playground free-generation window through November 23, 2026; API calls are billed. Confirm the live offer before you plan around it |
MiniMax's launch post does not publish a dollar price list in the blog text. It says hosted pricing is 50–70% of comparable mainstream offerings, with 50% off the standard price during launch week for requests at 4K and above. Check the live pricing page rather than a screenshot from launch day.
License limits you should actually read
Weights are open, not “do anything, anywhere.” The Community License (date November 10, 2026) is territorial and comes with an Acceptable Use Policy.
Key points as written, not legal advice:
- The Applicable Territory is worldwide except the European Union, the United Kingdom, the Republic of Korea, and the United States. MiniMax says people in those excluded territories should contact the company about a separate license.
- If your (and your affiliates') annualized revenue exceeds USD 1 million, or you offer a Hosted Service of the model to third parties, you must request prior written authorization.
- Use must follow the Acceptable Use Policy (Exhibit A). That policy includes a ban on circumventing or bypassing safety guardrails, on exploiting minors, and on a list of other prohibited harms. Hosted-service operators are also told to keep safeguards in place.
- MiniMax claims no rights over the outputs you generate; you remain responsible for them.
The optional PE model is noted as derived from Qwen3.8-27B (Apache 2.0) but is still subject to the H3-Image Community License.
If you do not want to self-host
A full local setup is a large download: the BF16 image checkpoint plus a separate 66.7 GB text encoder, and enough GPU (or a multi-GPU server) to run them. Official and third-party APIs remove that ops work, at the cost of their own pricing, rate limits, and content-safety layers.
If your goal is simply to generate or edit stills in a browser, without standing up H3-Image yourself, you can use a hosted workspace such as HackAIGC's uncensored image generator. That is a different stack, not a relabel of MiniMax H3-Image.
For the filter / adult-content question, use the companion page: Can MiniMax H3-Image Do NSFW?.
FAQ
Is MiniMax H3-Image the same as MiniMax H3?
No. H3 is the earlier H3-family video model. H3-Image is the still-image model built on that family, with open image weights of its own. See the H3 video guide if that is the product you meant.
Does it really output native 4K?
That is what MiniMax and the model card document: about 16 MP at common aspect ratios, including 4096 × 4096, generated directly rather than upscaled. The card's recommended working range is roughly 4–20 MP.
How big are the open weights?
The image checkpoint is listed at 68.65 GB (BF16 DiT). You also need the MiniMax-H3 text encoder (66.7 GB) unless you already have it. The optional prompt expander is 17.9 GB.
Can I run it in ComfyUI or on fal?
Yes, according to those companies' own launch notes: native ComfyUI support on November 10, 2026, and fal endpoints for generation and editing. Always check the live product; offers and templates change.
Can MiniMax H3-Image do NSFW?
That is a different question, and this explainer does not answer it with a lab test. Read Can MiniMax H3-Image Do NSFW?.
Bottom line
MiniMax H3-Image (announced November 10–11, 2026) is MiniMax's open-weights 4K-class image model: one checkpoint for text-to-image and multi-reference editing, documented at about 16 MP without an upscaler, distilled to 6 sampling steps, with official access via Hugging Face, the MiniMax API (including a Highspeed variant), Hailuo, MiniMax Code, plus day-one ComfyUI and fal routes. Speed and “best open-weight prompt adherence” figures are MiniMax's claims. The Community License is territorially restricted and is not a blank check. For adult-content and filter behavior, use the NSFW companion article. For a no-GPU browser path, try the uncensored image generator.
— Ethan Cole