NeoMME: a single-tower multimodal multilingual encoder
Oleh Zai · Let's Make It Easy
Dipublikasikan 2026-09-04
The interesting bit is architectural: document retrieval can use the page image directly, keeping layout, tables, and charts instead of depending on an OCR-f.
The interesting bit is architectural: document retrieval can use the page image directly, keeping layout, tables, and charts instead of depending on an OCR-first pipeline.
Apa yang dilaporkan
NeoMME is a 260M and 800M multilingual multimodal encoder that processes text tokens and raw image patches in one bidirectional Transformer.
NeoMME-Retriever reports 0.523 and 0.556 nDCG@10 on ViDoRe v3 for the 260M and 800M variants, respectively.
Detail penting
At matched 2048x2048 input on an NVIDIA L40S, the 260M model is reported to encode about 51 pages per second; the authors also report 255x smaller late-interaction indexes while retaining over 95% of baseline nDCG@10.
Catatan dan batas
The speed and retrieval figures are reported by the authors and depend on model size, hardware, input size, and benchmark setup. They are not a guarantee for every deployment.