MiniMax Releases Open-Weight Image Model H3-Image: Native 16-Megapixel 4K Images in About 10 Seconds, Up to 16 Reference Images, and a 33.1B DiT Distilled to 6 Steps

Core Background and Industry Pain Points

Most image models today output 1 to 4 megapixels, so print, poster, or large-screen work usually needs a separate upscaler, which is slow and can lose detail. Open models have also tended to trail closed products on instruction following and multi-reference editing. On November 10, MiniMax released H3-Image and published the weights on Hugging Face and GitHub at the same time. The community had already seen Pruna's P-H3-Image, which adapted the H3 video model; this is MiniMax's own official version.

Architecture Highlights and Underlying Mechanisms

H3-Image follows the H3-Base design with the audio branch removed. The prompt and reference images go through the Qwen3-VL text encoder from MiniMax-H3, reference images also pass through the image VAE, and all tokens form one packed sequence for a single-stream transformer. The diffusion transformer has 33.1B parameters in 50 dense layers. The VAE downsamples 16x spatially and each transformer token covers 32x32 pixels, which is what lets it produce 16MP images directly. The checkpoint is distilled to 6 sampling steps with no classifier-free guidance, which is the main reason 4K generation is fast. Generation and editing share a single checkpoint.

Authoritative Benchmarks and Measured Comparisons

MiniMax says a 4K image takes about 10 seconds and calls its instruction following the strongest among open-source models, but the model card gives no benchmark scores, so quantitative results have not been published and there is no independent evaluation yet. The speed figure describes MiniMax's hosted service; local times depend on your hardware. The official vLLM-Omni recipe lists deployment configurations for hardware such as the H200, RTX PRO 6000 Blackwell, and AMD Instinct MI350X.

Developer Adoption and Getting Started

The easiest path is the MiniMax API: text-to-image uses /v1/images/generations and editing uses /v1/images/edits, both compatible with the OpenAI SDK, and a faster MiniMax-H3-Image-Highspeed variant takes the same parameters. Based on the pricing page examples, a 4K text-to-image call costs about $0.069, with 50% off outputs of 16.6 megapixels or more until November 17. For local deployment, plan for about 68.65GB of DiT weights plus about 66.7GB for the text encoder; SGLang or vLLM-Omni can serve it on one GPU, and sequence parallelism across several GPUs lowers latency. Read the community license carefully before any commercial use.