NeoMME: A Single-Tower Multimodal-Native Multilingual Foundation Encoder for Efficient Fine-Tuning and Inference
NeoMME is pitched as a smaller encoder built specifically for multilingual text and raw image patches, not a repurposed generative VLM.
The paper introduces 260M- and 800M-parameter bidirectional encoder models with a 16,384-token context. Fine-tuned as retrievers, the 260M model reports 0.523 nDCG@10 on ViDoRe v3, while the 800M version reaches 0.556. The authors say the 260M model encodes 2048x2048 pages at about twice the throughput of ColModernVBERT on an NVIDIA L40S. They also report 255x compression for late-interaction multimodal document embeddings while retaining more than 95% of baseline nDCG@10. HF Daily Papers' note
The paper introduces 260M- and 800M-parameter bidirectional encoder models with a 16,384-token context. Fine-tuned as retrievers, the 260M model reports 0.523 nDCG@10 on ViDoRe v3, while the 800M version reaches 0.556. The authors say the 260M model encodes 2048x2048 pages at about twice the throughput of ColModernVBERT on an NVIDIA L40S. They also report 255x compression for late-interaction multimodal document embeddings while retaining more than 95% of baseline nDCG@10. HF Daily Papers' note
score 5