Pruna AI opens P-H3-Image, a smaller, faster image model built from MiniMax's H3 video model
Model compression company Pruna AI released open weights for P-H3-Image on November 3, an image generation and editing model built on H3, the video model MiniMax opened this summer. Pruna says H3's single images come out blurry and the model is big and slow, so it reworked each part: the text encoder distilled from 32B to 4B parameters, steps cut from 50 to 8, a new single-frame decoder, and 8-bit and 4-bit versions. In Pruna's own tests the smallest set is 20 GB versus 143 GB, a 4MP image takes 2.6 s on one NVIDIA B200 versus 12.4 s, and the 4-bit version needs about 21 GB of GPU memory at 8MP. Its MiniMax H3 Community License excludes the EU, UK, South Korea and US without a separate MiniMax license; hands, small background figures and small text can still break.
Why it matters
People who want a high-quality open image model on their own GPU get another option, though the license's regional limits need checking before any use.
Sources1
Written by dalao.ai: the title, summary and why it matters are our own words. The original reports are the record; their publishers keep authorship.