BLIP3-o

BLIP3-o is a family of fully open unified multimodal models released in May 2025, combining image understanding and image generation in a single architecture. It pairs an autoregressive backbone, strong at reasoning and instruction following, with a diffusion transformer that generates semantically rich CLIP image features rather than pixels, trained with a flow matching objective.

A sequential training strategy freezes the understanding backbone before training the generation side, preserving comprehension while adding strong generative ability. The release includes open weights, training code, and datasets, making it one of the most transparent unified multimodal efforts. A follow up, BLIP3o-NEXT, later pushed native image generation and editing further.