Pruna P-H3-Image: fast image model, but not for the EU

What happened?

On November 3, model compression company Pruna AI released P-H3-Image, an image generation model (Pruna). It is based on the MiniMax H3 video model, which Pruna turned into an image model that is 7x smaller and 5x faster. It generates images from 2 to 16 megapixels and can edit using several input images (the blog says 9, the Hugging Face page says 8).

How it got so small

• The full model shrank from 143 GB to 20 GB.
• Sampling steps dropped from 50 to 8 through DMD distillation.
• The text encoder was swapped from 32B to 4B parameters (9 GB instead of 67 GB).
• The backbone went from 33B to 20B parameters and got a new single-frame decoder.
• A prompt upsampler based on Qwen3.5-9B expands prompts.

Speed

According to Pruna, a 4-megapixel image takes 2.6 seconds on a B200 (versus 12.4 seconds originally) and an 8-megapixel image 5.8 seconds (versus 30.6). The 4-bit version takes 6.3 seconds on an L40S and needs about 21 GB of VRAM at 8 megapixels. The model runs natively in Diffusers and ComfyUI. Pruna itself notes that hands and small text remain weak spots.

The license is the catch

The weights are on Hugging Face, but the model falls under the MiniMax H3 Community License. It excludes the European Union, the United Kingdom, South Korea and the United States, where use requires a separate MiniMax license. Users must also display "Powered by MiniMax H3".

Why it matters

Technically it is an impressive result: a video model turned into a fast image model that fits on one GPU. For users in the EU, however, the license means it is not a freely usable model for now. Anyone planning production use needs a deal with MiniMax first.