MiniMax has released an image model, MiniMax H3-Image, and published the weights on Hugging Face at the same time. Text-to-image, image-reference generation and image editing are handled by one model. The company says it can produce a 4K image of 4096×4096, roughly 16 million pixels, in about 10 seconds, and accepts up to 16 reference images as input.
The model is available on Hugging Face and through the MiniMax open platform API, and has been wired into MiniMax's own products such as MiniMax Code and Hailuo. Partner platforms PixPix and PixTV can already call it directly, so users who don't want to host the model themselves can try 4K generation in a browser.
A few details from the model card
The backbone of H3-Image is a 33.1B-parameter diffusion Transformer with 50 layers and 56 attention heads, with weights in BF16. The text encoder is the Qwen3-VL encoder from the H3 video model, used unchanged; the DiT reads its hidden states from layer 50.
The speed comes mainly from distillation: the checkpoint is compressed to 6 sampling steps and runs without classifier-free guidance (CFG). Conventional diffusion models often need 20 to 30 steps, and with CFG enabled every step requires two forward passes. Removing both naturally cuts the time per image.
"The checkpoint is distilled to 6 sampling steps (6 NFEs) with no classifier-free guidance."
On resolution, the model card gives an effective range of 4 to 20 megapixels, with the best results around 16 megapixels. A 1:1 frame maps to 4096×4096 and 21:9 to 6272×2688, and neither needs a separate upscaling model. MiniMax also offers a Highspeed variant; it says standard mode is about 1.5 to 2 times faster than mainstream tools and high-speed mode about 3 to 4 times faster, without naming which tools it compared against.
Also open-sourced with the model is H3-Image-PE, a 27B prompt-expansion model fine-tuned from Qwen3.8-27B. It turns a short request into a detailed scene description and recommends a suitable aspect ratio. For image editing, only a prompt template has been released so far; the repository marks that part as still in development.
From video to images: how the H3 line got here
In late July, MiniMax first released the H3 video model and then opened its weights; SGLang later added inference acceleration for H3. H3-Image shares the same text encoder as the video model, while the image VAE is new: a CNN encoder paired with a 24-layer 2D ViT decoder, with 16x spatial downsampling, so each token covers 32×32 pixels.
Taken together, MiniMax's approach to multimodal generation has been consistent: ship video first, split out images, release the weights, and adapt the deployment frameworks in step. On launch day H3-Image came with four deployment paths, SGLang, vLLM-Omni, Diffusers and ComfyUI, and the weights are offered on both Hugging Face and ModelScope.
On instruction following, the official blog calls H3-Image the strongest among open-source models, but the model card includes no specific benchmark scores. Until third-party evaluations arrive, that claim rests on MiniMax's own account.
Who is likely to adopt it first
The 16-reference-image limit is practical for e-commerce and brand design: multi-angle photos of the same product, a brand color palette and model poses can all be fed in at once so the model keeps the subject consistent. One of the official showcase cases used 10 reference images.
Short-video creators can treat it as a step before video: lock in a 4K still frame and a character look with H3-Image, then hand it to the H3 video model to animate. PixTV's infinite canvas puts image and video generation on one canvas, so that whole chain can be completed in the browser; for product shots and detail pages, the e-commerce workspace in PixPix is a better fit.
On licensing, H3-Image uses the MiniMax-H3-Image community license, and some uses require written authorization through an official application form. Before commercial use, it is worth reading Section 4 of the agreement.
Sources: MiniMax official blog, Hugging Face model card, CocoLoop, MiniMax-H3-Image-PE model card; parameter count, sampling steps, resolution range and reference-image limit were checked, while speed multiples reflect MiniMax's own figures.