NanoVDR · DistilVDR · ColNanoVDR
NanoVDR is a family of small encoders aligned to the embedding space of a frozen vision-language retriever. A text encoder of 69M to 395M parameters answers each query, so pages indexed by a 2B to 8B model are searched without running that model per request. DistilVDR adds a 457M page encoder, which takes the large model out of indexing as well.
51ms
per query on one CPU thread for the 69M DistilBERT query encoder, at 95.1% of its 2B teacher on ViDoRe.
NanoVDR527M
parameters in the full DistilVDR pair. No multi-billion model runs at indexing or at query time.
DistilVDR95.7–98.1%
retention for multi-vector query encoders across five late-interaction teachers, trained without a single page image.
ColNanoVDROverview
Visual document retrieval matches a text query against page images, with no OCR in between. The strongest retrievers are 2B to 8B vision-language models. Indexing a corpus with one is a one-time cost; running the same model on every query is a cost that never stops. Each NanoVDR paper moves a different part of that work onto a small student.
Grey blocks are multi-billion vision-language models. Tinted blocks are NanoVDR students.
Each student minimises a direct discrepancy to cached teacher embeddings. No relevance labels, no hard-negative mining, no contrastive sampling.
An encoder only sees its own cached targets, so a query encoder and a page encoder can be trained separately and paired afterwards.
A student lands in its teacher's space, so it scores an index the teacher already built. Teacher and output width must match.
Models
Each paper replaces a different part of a vision-language retriever with a small student. All checkpoints are on the Hugging Face Hub under nanovdr.
Single-vector query encoders
A 69M to 151M text encoder replaces Qwen3-VL-Embedding-2B at query time. Pages stay in the teacher's index and scoring is a dot product.
Query and page encoders
Adds NanoVDR-D, a 457M page encoder, so both sides are students. A deployed pair has no multi-billion model in it at any point.
Multi-vector query encoders
A text encoder emits a set of token vectors in a ColPali-style teacher's space, so the existing page index is scored by MaxSim unchanged. Training is document-free.
Results
Every model with ViDoRe v3 results in the MTEB results repository, plotted by size against quality. NanoVDR and ColNanoVDR sit on the Pareto front: no listed model is both smaller and better. Each dashed arrow runs from a teacher to the student distilled from it.
Scroll sideways to see the whole chart.
ViDoRe v3, NDCG@5 averaged over the eight public tasks and their six query languages, computed from the MTEB results repository (snapshot of 2026-10-26, commit f530b79). Horizontal axis: parameters as listed on MTEB, which for NanoVDR and ColNanoVDR counts the query encoder; their pages are indexed by the teacher at the other end of the dashed line. DistilVDR is drawn at its full size, page and query encoder together. Its score comes from the DistilVDR paper under the same protocol; it has not been submitted to MTEB and the two private tasks were not run. 13 models below 20 NDCG@5 are not drawn: 10 CLIP and SigLIP variants, and bimodernvbert, modernvbert-embed and colSmol-256M. Hover or tap a point for its name.
| Model | Params on MTEB | Pages indexed by | NDCG@5, 8 public tasks | MTEB leaderboard |
|---|---|---|---|---|
NanoVDR-Q-DistilBERT-Qwen3VL2B-2048-ML | 69M | Qwen3-VL-Embedding-2B | 46.46 | 49.08 |
ColNanoVDR-Q-Ettin150M-ColVec8B-640-ML | 150M | webAI-ColVec1.1-8b | 60.18 | 62.35 |
ColNanoVDR-Q-Ettin150M-ColVec4B-640-ML | 150M | webAI-ColVec1.1-4b | 59.04 | 61.19 |
NanoVDR-D-HiRes-Qwen3VL8B-4096 +NanoVDR-Q-DistilBERT-Qwen3VL8B-4096-ML | 527M, not on MTEB | its own 457M page encoder | 47.07 (paper) | not submitted |
MTEB leaderboard: the score shown on the ViDoRe(v3) leaderboard, NDCG@10 averaged over all ten tasks including the two private ones. On MTEB, NanoVDR-Q-DistilBERT-Qwen3VL2B-2048-ML is still listed under its earlier name, NanoVDR-S-Multi.
Deployment
A query encoder and a page index work together when both target the same teacher at the same output width. The model name says which.
| Setup | Index pages with | Answer queries with | Multi-billion model at serving |
|---|---|---|---|
| NanoVDR | Qwen3-VL-Embedding-2B | NanoVDR-Q-*-Qwen3VL2B-2048 | indexing only |
| DistilVDR | NanoVDR-D-*-Qwen3VL8B-4096 | NanoVDR-Q-DistilBERT-Qwen3VL8B-4096-ML | never |
| ColNanoVDR | the matching ColPali-style teacher | ColNanoVDR-Q-* for that teacher | indexing only |
[Col]NanoVDR-<Q|D>-<variant>-<teacher>-<width>[-ML]
Col marks a multi-vector model. Q or D says which side it replaces. Teacher and width fix the embedding space; -ML means it was trained on the multilingual mixture.
Quick start
Every checkpoint loads with sentence-transformers and has no dependency on the NanoVDR repository.
from sentence_transformers import SentenceTransformer
query = SentenceTransformer("nanovdr/NanoVDR-Q-DistilBERT-Qwen3VL2B-2048-ML")
q_emb = query.encode(["What was the revenue growth in Q3 2024?"]) # (1, 2048)
scores = q_emb @ doc_emb.T # doc_emb: pages indexed by Qwen3-VL-Embedding-2Bfrom sentence_transformers import SentenceTransformer # >= 5.4 for the page encoder
query = SentenceTransformer("nanovdr/NanoVDR-Q-DistilBERT-Qwen3VL8B-4096-ML")
doc = SentenceTransformer("nanovdr/NanoVDR-D-HiRes-Qwen3VL8B-4096", trust_remote_code=True)
scores = query.encode(["..."]) @ doc.encode(pages).T # pages: PIL imagesfrom sentence_transformers import MultiVectorEncoder # >= 6.0
query = MultiVectorEncoder("nanovdr/ColNanoVDR-Q-Ettin150M-ColQwen35-320-ML",
trust_remote_code=True)
q = query.encode_query(["What was the revenue growth in Q3 2024?"])
scores = query.similarity(q, page_embeddings) # meanMaxSim, pages from the teacherPass the raw query. Do not add an instruction prefix. The teacher's instruction was applied when its targets were cached, and every published number was measured on bare query text.
FAQ
The query side can. The 69M NanoVDR query encoder answers a query in 51 ms on one CPU thread. Pages are still encoded once, offline, on a GPU, either by the teacher or by a NanoVDR-D page encoder.
Yes, if one of the supported teachers built it: ColQwen3.5-4.5B, Vultron-4.5B, Tomoro-8B, ColVec1.1 4B or 8B, or EVIE 4.5B or 8B. The matching ColNanoVDR query encoder emits token vectors in that teacher's space, and the index is scored by MaxSim unchanged.
No. Every encoder is trained against cached teacher embeddings with a single alignment loss. ColNanoVDR also needs no page images: it only consumes the teacher's cached query token embeddings.
The -ML models were trained on English plus five Latin-script European languages, translated with MarianMT. Non-Latin scripts are untested.
For a single-vector index built by Qwen3-VL-Embedding-2B, NanoVDR-Q-DistilBERT-Qwen3VL2B-2048-ML. For a deployment with no large model at all, the DistilVDR pair. For a ColQwen3.5 multi-vector index, ColNanoVDR-Q-Ettin150M-ColQwen35-320-ML.
Papers
Zhuchenyang Liu, Yao Zhang, Yu Xiao
@inproceedings{liu-etal-2026-nanovdr,
title = "{N}ano{VDR}: Distilling a 2{B} Vision-Language Retriever into a 70{M} Text-Only Encoder for Visual Document Retrieval",
author = "Liu, Zhuchenyang and
Zhang, Yao and
Xiao, Yu",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.71/",
pages = "1631--1648",
ISBN = "979-8-89176-459-0"
}Zhuchenyang Liu, Ziyi Wang, Yao Zhang, Yu Xiao
@inproceedings{liu-etal-2026-distilvdr,
title = "{D}istil{VDR}: A Compact End-to-End Visual Document Retriever via Dual-Student Distillation",
author = "Liu, Zhuchenyang and
Wang, Ziyi and
Zhang, Yao and
Xiao, Yu",
editor = "T.Y.S.S., Santosh and
Agarwal, Amit and
Gupta, Vivek and
Kementchedjhieva, Yova and
Garimella, Aparna and
Hou, Yufang",
booktitle = "The Proceedings of Workshop on Document Intelligence and Understanding",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.docinsights-1.26/",
pages = "385--399",
ISBN = "979-8-89176-477-4"
}Zhuchenyang Liu, Ziyi Wang, Yao Zhang, Yu Xiao
@misc{liu2026colnanovdrdocumentfreequerydistillation,
title={ColNanoVDR: Document-Free Query Distillation for Multi-Vector Visual Document Retrieval via Optimal Transport},
author={Zhuchenyang Liu and Ziyi Wang and Yao Zhang and Yu Xiao},
year={2026},
eprint={2609.34899},
archivePrefix={arXiv},
primaryClass={cs.IR},
url={https://arxiv.org/abs/2609.34899},
}