MiniMax releases open-weight H3-Image for 4K image making and editing

MiniMax has published open weights for MiniMax-H3-Image, a single model that both creates images from text and edits existing images, producing pictures of about 16 megapixels directly without an upscaler. The release, which drew attention on X on November 11, works in ComfyUI from day one, but its licence requires written permission from MiniMax for larger companies and hosted services.

The model comes from the same family as MiniMax H3, the open-weight video model the company released in early August. Where H3’s open checkpoint was limited to 768p video and kept its 2K stage behind the API, H3-Image’s downloadable checkpoint produces its full output resolution natively, according to MiniMax’s model card.

What MiniMax H3-Image does

MiniMax describes H3-Image as a model for image generation and editing in which “generation and editing are unified in one checkpoint.” It generates images from text and edits one or more reference images. A checkpoint is the saved set of trained weights that developers download and run.

The headline specification is output size. The model card says H3-Image generates images of about 16 megapixels at every aspect ratio, from 4096x4096 for a square (1:1) image to 6272x2688 for an ultra-wide 21:9 image, “directly and without an upscaler”. MiniMax contrasts this with “the 1–4 MP outputs common today” and says 16 MP images “typically require substantially more computation,” which the company says it keeps manageable.

Speed comes from distillation, a training technique that compresses a model’s sampling process into fewer steps. The checkpoint is distilled to 6 sampling steps (6 NFEs, or function evaluations) and runs without classifier-free guidance, a common technique that roughly doubles the work per step. For editing, the model accepts up to 16 reference images.

Spec	MiniMax-H3-Image (per MiniMax)
Tasks	Text-to-image and image editing, one checkpoint
Output size	About 16 MP; 4096x4096 (1:1) to 6272x2688 (21:9)
Upscaler	None needed
Sampling steps	6, no classifier-free guidance
Reference images	Up to 16
Diffusion transformer	33.1B parameters, 50 layers, BF16 weights
Text encoder	Qwen3-VL text encoder from MiniMax-H3, unchanged
Main repository size	68.65 GB (BF16 DiT listing)
Optional prompt expander	MiniMax-H3-Image-PE, NVFP4, 17.9 GB

Source: MiniMax model card on Hugging Face.

How H3-Image differs from the H3 video model

H3-Image reuses most of the H3 stack. Its diffusion transformer (DiT, the network that turns noise into an image step by step) “has the H3 architecture without the audio layers,” and it uses the MiniMax-H3 text encoder unchanged. The Comfy-Org repository lists MiniMax-H3 as the base model and H3-Image as a fine-tune of it.

The shared parts are visible in the numbers. MiniMax gives the H3-Image transformer 33.1B parameters across 50 layers, with a hidden size of 5376 and 56 attention heads, “the same as the H3-Omni-Transformer”. Coverage of the H3 video model described the same 50-layer, 5,376-hidden-size, 56-head design at 33 billion parameters, with Qwen3-VL-32B as the text encoder.

The main practical difference for builders is what ships open. When H3 launched, MiniMax described a three-part system in which a context-processing layer and a 2K regeneration stage were not included in the open-weight release. Oxen.ai noted that the open release contained only the “base generation” block, so local users had to build their own pipeline to match the API. H3-Image, by contrast, is presented as producing its 16 MP output from the published checkpoint alone.

Architecture in brief

• Inputs: the text prompt and reference images pass through the H3 text encoder; reference images are also encoded by the image VAE (the component that compresses images into a smaller latent form). All tokens form one packed sequence for a single-stream transformer.
• Text encoder: the full Qwen3-VL text encoder of MiniMax-H3; the DiT reads hidden states from its 50th layer.
• Image VAE: a CNN encoder and a 24-layer 2D ViT decoder, with 16x spatial downsampling and 24 latent channels. With a 2x2 patch, each transformer token covers 32x32 pixels.
• Precision: BF16 weights, with the embedding, timestep and output layers in FP32.

Where H3-Image is available and how to run it

MiniMax offers the model three ways. It is available through the company’s API on separate global and China platforms, with documentation for text-to-image, image editing and prompt expansion. It is also available in MiniMax’s Hailuo web app and its desktop agent app. The weights are on Hugging Face.

For local deployment, MiniMax recommends four frameworks: SGLang, vLLM-Omni, diffusers and ComfyUI, with separate ComfyUI templates for text-to-image and editing. The model card says both NVIDIA and AMD accelerators are supported, “including NVIDIA H200, RTX PRO 6000 Blackwell, AMD Instinct MI350X and more”. A single GPU can serve the model, and for lower latency MiniMax shows how to split each request across several GPUs with sequence parallelism, giving a four-GPU example.

Download sizes are large. The transformer folder alone is 66.28 GB and the VAE is 2.34 GB. The text encoder is not in the H3-Image repository; it loads from the MiniMax-H3 repository and weighs 66.7 GB, though it is not downloaded again for anyone who already has MiniMax-H3. The model card does not list a minimum VRAM figure for H3-Image.

Comfy-Org, the organisation behind ComfyUI, hosts repackaged files for the node-based tool. They include a pruned int8 diffusion model, int8 and NVFP4 versions of the Qwen3-VL text encoder and an FP16 VAE. For prompt enhancement inside ComfyUI, Comfy-Org recommends a separate quantised Qwen3.5 9B model with its Generate Text node.

The licence: tighter than H3’s

The weights are downloadable, but the terms matter more than usual. The licence note on the Comfy-Org repository says the H3-Image licence agreement “is different from the original MiniMax H3 license”. Under it, users need “a separate, prior written authorization from MiniMax” if the annualized revenue of the user and their affiliates exceeds $1 million, or if they provide hosted services to third parties.

That is a much lower bar than for H3. MiniMax’s H3 FAQ says the H3 Community License grants royalty-free commercial use, with separate authorization required only above US$20 million in yearly revenue. The H3 terms also required products to display “MiniMax H3” in their interface and banned using outputs or weights to train other AI models. The sources do not show whether those clauses carry over to H3-Image.

On geography, a post on X that flagged the release says the H3-Image licence “excludes the US, EU, UK and South Korea”. The sources here do not include the H3-Image licence text itself, so that point rests on the post. It would match H3, whose community licence applies worldwide except in those four territories.

MiniMax has explained the H3 exclusions before. For the US, MiniMax’s developer-relations lead Ryan Lee linked the restriction to the company’s copyright lawsuit with Disney, Universal and Warner Bros. Discovery over its Hailuo video service, according to we0.ai. For the EU, UK and South Korea, MiniMax pointed to changing regulation around synthetic media, likeness and copyright. MiniMax’s own H3 FAQ calls the exclusion “not yet,” not “not ever,” and says organisations in excluded regions can apply for a formal licence while the API remains available globally.

MiniMax’s claims versus independent evidence

The quality claims come from MiniMax. The model card calls H3-Image “state-of-the-art” and says it “delivers the strongest instruction following among open-source models”. It also says the model “goes beyond the familiar ‘AI look’” to produce realistic detail. The model card text in the sources includes no benchmark table to back these statements, and none of the sources report a third-party image leaderboard result for H3-Image.

The parent model did get outside evaluation. Spheron, citing Artificial Analysis, reported that H3 ranked first in video editing, second in text-to-video and third in image-to-video with audio among evaluated models as of early August 2026, calling it “a single third-party arena run”. That says nothing direct about still-image quality, and builders should treat H3-Image’s instruction-following claim as unverified until independent tests appear.

What it means for builders

For developers who self-host, H3-Image is unusual in pairing open weights with native 16 MP output and six-step sampling in one checkpoint that both generates and edits. Teams that ran H3 locally may already have the 66.7 GB text encoder, which cuts the extra download. Support across SGLang, vLLM-Omni, diffusers and ComfyUI means it can slot into existing serving stacks or node workflows.

The limits are mostly legal and hardware-related:

• Hosted products need permission. Anyone offering H3-Image as a hosted service to third parties needs prior written authorization, regardless of revenue. That rules out quietly wrapping it in a public app or API.
• The revenue threshold is low. At $1 million in annualized revenue including affiliates, many small startups would need authorization for use above the threshold.
• Regional exclusion is reported. Teams in the US, EU, UK or South Korea should check the licence text before downloading, given the X report.
• Heavy weights. The full-precision transformer and text encoder together exceed 130 GB on disk, based on MiniMax’s listed sizes.

Builders who want the model without licence questions about self-hosting have MiniMax’s own API and apps as the official route.

What to watch next

Three things will decide how widely H3-Image gets used. First, independent tests of its instruction following and 16 MP detail against other open image models, since current claims are MiniMax’s own. Second, whether MiniMax publishes the full H3-Image licence alongside the weights and widens its territory scope, as it has said it plans to do for H3 when the regulatory picture clarifies. Third, community work on smaller builds: Comfy-Org already ships int8 and NVFP4 variants, and H3’s release saw rapid community optimisation within its first week.

For more on open generative media models, see Unicorn API’s coverage of Lightricks’ LTX-2.5 open video model.