Chinese AI firm MiniMax has released H3-Image, an open-weight image model that both generates and edits pictures at resolutions reaching 4K. The model went live on November 10, 2026.

H3-Image is available on Hugging Face—a widely used platform for sharing open machine-learning models—under the MiniMaxAI organization, through MiniMax Code, and via the company's APIs. Developers can therefore download the weights and run the model on their own infrastructure, or pay per image through the hosted API instead.

There is a catch for a large share of the world's developers, however. The model's community license excludes use in the United States, the European Union, the United Kingdom, and South Korea. The weights are public, but the welcome mat has several major markets crossed off.

One Model for Generation and Editing

Most image tools divide the work between two systems: one model creates images from text, while a separate model handles edits. H3-Image consolidates both jobs into a single framework, covering text-to-image generation and advanced editing within one architecture.

Under the hood, the model runs on a diffusion transformer (DiT) architecture, a design that pairs diffusion-based image generation with transformer components—the architecture family underpinning most modern language models—and carries 33.1 billion parameters, the internal settings the system learned during training.

Native 4K Resolution

Output resolution is the headline feature. H3-Image produces native images of approximately 16 megapixels, with supported sizes including 4096×4096. It can also generate wider formats up to 6272×2688, an aspect ratio better suited to panoramas and banners than to profile pictures. For context, a standard 4K display resolves roughly 8.3 megapixels, so the model's native output exceeds that figure at its larger supported sizes.

The word "native" carries weight here. Many image models generate at a lower resolution and then upscale the result—in effect stretching a smaller picture and hoping nobody zooms in. Native output means the model constructs the image at full size from the start.

Speed, Reference Images, and Local Requirements

Generation is efficient: H3-Image requires only about 6 sampling steps—the iterative refinement passes a diffusion model runs to arrive at a finished image—and it does so without relying on classifier-free guidance, a technique widely used to tighten how closely outputs follow their prompts, at the cost of extra computation.

The model also accepts up to 16 reference images at once. That gives users a way to show the model what they want—a specific character, product, or style—rather than describing it entirely in words.

API pricing runs approximately $0.022 for a 1K image and approximately $0.069 for a 4K image, with output size driving the cost. MiniMax is additionally offering promotional discounts until November 17, 2026.

Running the model locally carries an additional requirement. Full functionality depends on a separate text encoder, Qwen3-VL-32B, which translates prompts into a form the image model can process. The download is therefore effectively two large models, not one.

Integration arrived quickly: ComfyUI, the node-based interface widely used for building image generation workflows, supported H3-Image immediately upon its release.

Licensing Restrictions

The licensing question extends beyond geography. In addition to barring use across the US, EU, UK, and South Korea, the terms require larger companies to obtain written authorization from MiniMax before using the model. Open weight, in other words, does not mean open use: the weights are public, but the license terms still govern who can deploy them, and where.

A Derivative Already in the Mix

The ecosystem around H3-Image is already moving. A derivative model called P-H3-Image was announced on November 3, 2026, with claims of faster inference on consumer hardware. Building on published weights is exactly the kind of follow-on work open-weight releases enable.

Source: CryptoBriefing