NeoMME introduces a multimodal-native, multilingual encoder

NeoMME's 260M and 800M models process raw image patches and multilingual text in one bidirectional Transformer with a 16,384-token context. The authors report 0.556 nDCG@10 on ViDoRe v3 for the 800M model and release the pretrained backbone and retrieval-compatible checkpoints under Apache 2.0.

Visit original source