Pruna released P-H3-Image on November 3. It's an open-weights model for text-to-image and image editing, built from MiniMax-H3, the video model MiniMax opened this summer. Pruna's reasoning: a video model already has to learn how objects look from different angles and how light falls on them, so it's a good base for stills.
They changed every part of H3. A small prompt upsampler (based on Qwen3.5-9B) rewrites short requests into prompts H3 handles well. The 32B text encoder was distilled to 4B (9 GB instead of 67 GB). Sampling dropped from 50 steps to 8, a new decoder turns one frame into one sharp image, and the backbone went from 33B to 20B parameters.
Pruna's own numbers: a 4 MP image takes 2.6 seconds on one B200 against 12.4 for H3, and 6.3 seconds on an L40S with the 4-bit ComfyUI set. Downloads are 55 GB in bf16 and 20 GB in 4-bit, with peak VRAM at 8 MP of 56 GB and 21 GB. The bf16 Diffusers pipeline with the upsampler needs about 100 GB at 8 MP, so most people will want the ComfyUI weights. Pruna recommends 8 MP for quality and says small or dense text, hands and background figures can still come out wrong.
Editing takes several input images. Pruna's post says up to nine; The Model Atlas reads the model card as one to eight, with the first image edited and the rest used as references.
The license is the catch. P-H3-Image inherits the MiniMax H3 Community License, which, per Pruna, excludes the European Union, the United Kingdom, South Korea and the United States without a separate license from MiniMax. The Model Atlas also points to clauses requiring "MiniMax H3" in the interface of commercial products, written permission above $20M in yearly revenue, and no using outputs to improve other models.