From Pixels to Words -- Towards Native One-Vision Models at Scale

Researchers introduced NEO-ov, a native vision-language foundation model that learns pixel-word and cross-frame relationships end to end.

The model removes external encoders, adapters, and post-hoc fusion, aiming for a simpler multimodal architecture.

NEO-ov shows strong performance on fine-grained visual tasks and comes close to modular systems overall.

The paper also offers architecture and training guidance for building native vision-language models.