AI News
All stories
Reported

MiniMax releases open-weights H3 Image model, 68 GB or 18 GB quantized

Matthew Berman says MiniMax's H3 Image is a fully open-weights image model, 68 GB at full size and about 18 GB in a quantized version.

Models · · 2 min read

MiniMax has published H3 Image on Hugging Face as open weights. Its model card describes one checkpoint that both generates images from text and edits up to 16 reference images. Matthew Berman said the full model is 68 GB and that it can be downloaded now. The card's model zoo table lists the BF16 image model at 68.65 GB, and its download section lists the transformer folder at 66.28 GB. The card also says the text encoder is not in the repository. It is loaded from the separate MiniMax-H3 repository, at 66.7 GB, and is not downloaded again if the user already has MiniMax-H3. A first-time local setup therefore needs more than the 68 GB figure suggests.

The card describes a 33.1B-parameter dense diffusion transformer with 50 layers. It is distilled to six sampling steps with no classifier-free guidance. The image VAE uses 16x spatial downsampling, so each transformer token covers 32x32 pixels. MiniMax says the model outputs about 16 megapixels at every aspect ratio, from 4096x4096 up to 6272x2688 at 21:9, without an upscaler. The card says the model works best around 16 MP and gives a working range of about 4 MP to 20 MP. Producing images that size typically takes substantially more computation than the 1 to 4 MP outputs the card compares against. The text does not back up the claim that MiniMax keeps that cost manageable with latency figures. Instead it points to a vLLM-Omni recipe for measured latency per hardware configuration.

Berman's 18 GB figure does not match the repository. The 17.9 GB file in the model zoo is an NVFP4 build of the optional MiniMax-H3-Image-PE prompt expander for text-to-image. The image model itself is listed only at BF16. Berman also said the quantized version can run on most current computers. The card says nothing about consumer hardware. It recommends SGLang, vLLM-Omni, diffusers or ComfyUI. Its vLLM-Omni section names accelerators such as NVIDIA H200, RTX PRO 6000 Blackwell and AMD Instinct MI350X. Its diffusers example loads the model onto cuda, with a note to switch to mps on Apple devices.

Berman called H3 Image fully open-source open weights. The card says the model is released under the MiniMax-H3-Image Community License Agreement. It also says a separate written authorization, described in Section IV of that license, is requested through an application form.

The card makes its own performance claims. It says H3-Image delivers the strongest instruction following among open-source models and supports up to 16 reference images for more control. It also says the API turns a short prompt into a detailed English description with the H3-Image-Context-IR step before generation. The local workflow, the card says, reproduces the quality of API-generated images. The card offers the optional MiniMax-H3-Image-PE expander as a local alternative for that prompt step, and says the expander is subject to the same license agreement.

What changed

MiniMax published H3 Image on Hugging Face as open weights, a model that generates images from text and edits up to 16 reference images.

Type: New capability · Applies to: H3 Image, MiniMax-H3-Image-PE, MiniMax-H3

Why it matters

A downloadable image model that runs on consumer hardware lets teams keep image generation in-house and avoid per-image API fees.

Original source

Where we saw it

1 source

Forward Future (Matthew Berman)

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works