MiniMax Releases H3-Image, an Open-Weight Image Model That Generates 16 Megapixels in Six Steps

On November 10, 2026, MiniMax launched H3-Image, an open-weight image generation and editing model. It directly outputs images of around 16 megapixels (4096 × 4096 pixels in square format) without an upscaling stage, in just six computation steps, and accepts up to 16 reference images. It is also available via the MiniMax API, starting at around $0.022 per image. As with other models in the H3 family, its license excludes the European Union.

A Video Model Turned Image Generator

H3-Image is derived from MiniMax H3, the developer’s open video model. According to the Hugging Face model card, its diffusion transformer adopts the H3 architecture without the audio component: 33.1 billion parameters, 50 layers, and the same text encoder (based on Qwen3-VL), reused as is. The model was distilled for six sampling steps, without CFG guidance or negative prompts. Both generation and editing fit into a single weight file.

This is not the first attempt of its kind. A week ago, Pruna AI released P-H3-Image, a third-party derivative of H3 running in eight steps. MiniMax’s official release follows the same path but aims for higher resolution: the model card advertises an effective range of roughly 4 to 20 megapixels, with peak quality around 16 megapixels, whereas most current generators produce 1 to 4 megapixels.

The model expects a detailed image description in English. The API automatically writes this prompt from a short request in any language, and MiniMax also published a locally deployable prompt-rewriting model, H3-Image-PE.

Requirements to Run It

For local deployment, you need substantial storage: 66.28 GB for the transformer in BF16, 2.34 GB for the image decoder, and 66.7 GB for the H3 text encoder, which must be downloaded separately. MiniMax recommends SGLang, vLLM-Omni, diffusers, or ComfyUI, citing hardware such as the NVIDIA H200, RTX PRO 6000 Blackwell, or AMD Instinct MI350X. This is far beyond consumer GPUs, unlike Pruna’s 4-bit version, which fit into roughly 21 GB of VRAM.

The simplest option remains the API, available in standard and Highspeed tiers (faster and more expensive). The pricing table provides examples: around $0.022 for a 1024 × 1024-pixel image, $0.069 for 4096 × 4096 in the standard tier, and $0.13 in Highspeed. Until November 17 at 16:00 UTC, MiniMax is cutting the price in half for images of at least 16.6 million pixels. The model is also available within MiniMax Code and the Hailuo AI app. The API does not support inpainting masks or transparent backgrounds.

What Remains to Be Verified

MiniMax claims the model can generate a 4K image “in 10 seconds” and boasts superior instruction following compared to other open models. No independent benchmarks have been published so far, and the announcement does not specify the hardware used. The examples in the model card (realistic photographs, video game maps transformed into 3D views, and a fully furnished room generated from ten images) were selected by the developer.

A License Excluding Europe and the United States

The weights are not distributed under an open-source license. The MiniMax-H3-Image Community License excludes the European Union, the United Kingdom, the Republic of Korea, and the United States, unless a separate license is granted by MiniMax. It strictly forbids using, modifying, or distributing the weights—or their outputs—outside the authorized territory. Commercial products relying on it must display “MiniMax-H3-Image” on their interface, and written permission from MiniMax is required for businesses exceeding $1 million in annual revenue or looking to host the model as a third-party service. This license covers the weights; the API and Hailuo AI are governed by their respective terms of service.