Hcompany/NeoMME-260M-Retriever-ST-dense

NeoMME-Retriever (260M): Single-Tower Multimodal-Native Multilingual Foundation Encoder 🔎

NeoMME-Retriever (260M) variants:

• Default (transformers): Returns dense and multi-vector embeddings together with a single forward pass. Recommended for most use cases and inference.
• ST dense [current]: Supports independent dense fine-tuning with Sentence Transformers.
• ST late-interaction: Supports independent multi-vector fine-tuning with Sentence Transformers.

NeoMME-260M-Retriever-ST-dense is a model for multimodal document retrieval. Fine-tuned from NeoMME-260M, it encodes text queries and documents (text or page screenshots) using one shared bidirectional Transformer encoder.

This model can be used with Sentence Transformers, but can only generate dense embeddings.

Specification	Value
Parameters	263M
Vocabulary	131,072 tokens
Context length	16,384 tokens
Hidden size	1,024
Image patches	32 × 32 pixels, up to 2,048 pixels on the longest side (default)
Dense embeddings	1,024 dimensions (Matryoshka: [128, 256, 512, 1,024])
Dense pooling strategy	Mean

Dense embeddings are L2-normalized and use cosine similarity. They match NeoMMEForRetrieval.dense_embeddings.

Performance

All scores use the metric shown at the full trained dimensions. Higher is better. ViDoRe v3, v2, and v1 measure visual document retrieval, while BEIR-15 measures text retrieval.

Benchmark	Metric	NeoMME-260M		NeoMME-800M
		Late interaction	Dense [current]	Late interaction	Dense
ViDoRe v3	nDCG@10	0.5226	0.3907	0.5560	0.4391
ViDoRe v2	nDCG@5	0.5218	0.4075	0.5591	0.4475
ViDoRe v1	nDCG@5	0.8598	0.7552	0.8744	0.7993
BEIR-15	nDCG@10	0.4881	0.3055	0.5126	0.3686

Usage

pip install -U "sentence-transformers[image]"

from sentence_transformers import SentenceTransformer


model = SentenceTransformer("Hcompany/NeoMME-260M-Retriever-ST-dense")

queries = [
    "Quelle partie de la production pétrolière du Kazakhstan provient de champs en mer ?",
    "Which hour of the day had the highest overall electricity generation in 2019?",
]
documents = [
    "https://github.com/tonywu71/colpali-cookbooks/blob/main/examples/data/shift_kazakhstan.jpg?raw=true",
    "https://github.com/tonywu71/colpali-cookbooks/blob/main/examples/data/energy_electricity_generation.jpg?raw=true",
]

query_embeddings = model.encode_query(queries, convert_to_tensor=True)
document_embeddings = model.encode_document(documents, convert_to_tensor=True)
scores = model.similarity(query_embeddings, document_embeddings)

# Expected: scores[0, 0] > scores[0, 1] and scores[1, 1] > scores[1, 0].
print(scores)

The score tensor has shape (num_queries, num_documents) and scores[i, j] is the score between query i and document j. A larger value indicates a closer match.

Training

NeoMME-260M-Retriever was fine-tuned from NeoMME-260M on text retrieval and document-page images. Training uses a joint late-interaction and Matryoshka dense contrastive objective.

The NeoMME technical report describes the full fine-tuning recipe (will be released soon).

Limitations

With Sentence Transformers, only one of the two retrieval heads can be used at a time.

License

Model weights are released under the Apache 2.0 license.

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},
}