@Chinese AI company MiniMax launched the open-weight model 'H3-Image (H3-Image)' that can generate 4K images in about 10 seconds. It is characterized by the ability to handle image generation and editing with a single model and to utilize up to 16 reference images.
MiniMax announced the release of H3-Image on the 11th through an official statement, and stated that it has deployed the model weights and related code on Hugging Face and GitHub. It can also be used via MiniMax API and AI development tools 'MiniMax Code', as well as video and image generation services 'Hailuo AI'.
H3-Image is designed to directly generate images with a resolution of approximately 16 million pixels (4096×4096) without additional upscaling. MiniMax emphasized that it can create high-quality images up to five times faster than major commercial models.
According to the API latency comparison table released by the company, generating a 4K image takes 16 seconds in standard mode and 10.3 seconds in fast mode. Google's 'Nano Banana 2.1' takes 48 to 64 seconds for the same resolution.
However, it did not show a clear advantage in low-resolution scenarios. For 1K image generation, H3-Image's standard mode takes 10.6 seconds, which is slower than the mid-mode of 'GPT Image 2.6' (9 seconds).
@It supports image editing as well as image generation with the same model. You can combine elements from multiple images or create new results while maintaining the appearance of specific people and objects, by inputting up to 16 reference images. It is designed to perform various creative tasks without the need for separate editing models or additional adapters.
In terms of image quality, it emphasizes natural textures and detailed expressions. It aims to reduce the overly smooth and artificial feel often seen in existing AI-generated images, focusing on achieving results close to real photographs. It also highlights its ability to accurately reflect the details of prompts and the characteristics of reference images as a key strength.
4K image generation in 10 seconds! ⚡
Meet MiniMax H3-Image.
16 million pixel resolution. Stunning realism. Precise prompt following & editing. Up to 16 reference images. All in 10 seconds.
And it's open-weight.
Available now via our API, MiniMax Code, and Hailuo AI — with… pic.twitter.com/ZLprDpzpv5
— MiniMax (official) (@MiniMax_AI) November 10, 2026
This model was developed based on the multi-modal generative model 'H3' announced in July. Minimax explained that it improved subsequent learning processes such as data composition and reinforcement learning, and reduced the sampling steps to six, thereby increasing generation speed. It uses a diffusion transformer (DiT) structure with 3.31 billion parameters.
It also emphasized price competitiveness. The API cost is approximately $0.022 for generating one 1K image, $0.023 for 2K, and $0.069 for 4K. The high-speed mode increases the cost to about $0.130 based on 4K.
A 50% discount is applied to requests for generating images larger than 4K until the 17th. Minimax explained that even before the discount, the prices are at the level of 50 to 70% of major competing services.
The weights are public, but there are restrictions on usage conditions. South Korea, the United States, the United Kingdom, and the European Union (EU) are excluded from the basic community license area, so businesses or individuals wishing to use or deploy the model directly need separate permission. Even in permitted regions, businesses providing the service to external users or those with annual revenue exceeding $1 million must obtain prior written approval from Minimax.
With this release, Minimax expanded its product line to include high-resolution image generation, following the video and audio generation model H3. The company highlighted fast generation speed and high image quality, but independent benchmarks for prompt understanding and image quality have not yet been confirmed.
Park Chan, cpark@aitimes.com
