NEO

Why we included this project

NEO comes at multimodal modeling from an angle most VLM projects don't: instead of attaching a separate vision encoder to an LLM, it trains pixels and text together inside a single dense architecture. The repo backs that idea with real artifacts. 2B and 9B checkpoints on Hugging Face, the training recipes used to build them, and an evaluation harness let you reproduce the reported results or fine-tune a model for your own task rather than taking the papers on faith. The efficiency claim is what makes it interesting: the authors say visual understanding is comparable to modular systems while training on a fraction of the image-text data. Anyone weighing current open vision-language approaches, especially teams curious about encoder-free designs, will find working code here and a useful reference point.