Infrastructure & deployment
NeoMME: an efficient Multimodal-native and Multilingual Encoder
NeoMME uses one from-scratch bidirectional Transformer for text and image patches, then adapts it for visual document retrieval with dense and late-interaction outputs.
Why it matters: For visual document retrieval, ranking page images preserves layout and other visual clues without OCR preprocessing. The reported compression and encoding results suggest ways to manage index storage and encoding cost, though they do not establish end-to-end PDF processing or retrieval latency.