NEO: Native Vision-Language Models

NEO explores native vision-language models: pixels and words are processed together inside an encoder-free model, rather than connecting a separately pretrained vision encoder to a language model.

Code and model collections · Original paper

The NEO family

• NEO introduces native vision-language primitives for single-image understanding. Paper
• NEO-ov extends the approach to multi-image and video understanding. Paper
• NEO-unify explores end-to-end unified multimodal models, with subsequent work in the SenseNova-U1 family. Technical introduction
• NEO-VLA brings native vision-language representations to robot action policies. Read the project article

Pixels and words, together

The original architecture embeds image patches and text tokens into the same model. The shared backbone learns visual and language representations together.

The official repository maintains the family’s training and evaluation guides, model weights, datasets, and release history. The architecture figure and project artwork above come from that repository.