hugging-apps/neomme-260m-retriever-demo

NeoMME-260M-Retriever — late-interaction document search

Demo of `Hcompany/NeoMME-260M-Retriever-ST-late`: a 263M-parameter, multilingual, single-tower encoder that embeds a text query and document page images into the same 128-d multi-vector space and ranks them with MaxSim late interaction — no OCR and no text extraction, the model reads pixels.

Upload page images or PDFs, type a question, and the app returns the pages ordered by score plus a similarity map showing which 32×32 patches of the winning page matched the query. Plain-text passages can be scored alongside the pages (the same checkpoint tops BEIR-15 text retrieval for its size class).

The model is loaded through Sentence Transformers' MultiVectorEncoder, exactly as the model card documents; scoring uses MeanMaxSim.

Example assets

• —examples/shift_kazakhstan.jpg and examples/energy_electricity_generation.jpg — the two pages used in the NeoMME model card's usage snippet, from tonywu71/colpali-cookbooks (MIT, © 2024 Manuel Faysse, Hugues Sibille, Tony Wu).
• —The other six pages are sampled from the ViDoRe v1 benchmark suite this model reports results on — vidore/docvqa_test_subsampled, vidore/infovqa_test_subsampled, vidore/tabfquad_test_subsampled, vidore/syntheticDocQA_government_reports_test, vidore/syntheticDocQA_energy_test and vidore/shiftproject_test (all MIT).

Citation

@misc{lac2026neommesingletowermultimodalnativemultilingual,
      title={NeoMME: A Single-Tower Multimodal-Native Multilingual Foundation Encoder for Efficient Fine-Tuning and Inference},
      author={Aurélien Lac and Tony Wu},
      year={2026},
      eprint={2609.01657},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2609.01657},
}