Paper: NanoVDR: Distilling a 2B Vision-Language Retriever into a 70M Text-Only Encoder for Visual Document Retrieval | Blog | Project page

NanoVDR-Q-BERT-Qwen3VL2B-2048-ML: Multilingual Query Encoder for Visual Document Retrieval

⚠️ This model was renamed

It was published as nanovdr/NanoVDR-M-Multi. Old links and from_pretrained calls still resolve through a redirect, but please move to the new id.

- SentenceTransformer("nanovdr/NanoVDR-M-Multi")
+ SentenceTransformer("nanovdr/NanoVDR-Q-BERT-Qwen3VL2B-2048-ML")

Why. NanoVDR started as one thing: a small text encoder that replaces the query side of a large vision-language retriever. It has since grown a second tower that replaces the document side, and a second teacher, so a name like -S no longer says enough. Two checkpoints only work together when they were distilled from the same teacher into the same width, and neither fact was recoverable from the old names.

The scheme is now NanoVDR-<Q|D>-<variant>-<teacher>-<width>[-ML], and the rule is simply that the teacher and the width have to match. This model is Q (query tower), BERT (backbone), distilled from Qwen3-VL-Embedding-2B into 2048 dimensions, trained on the multilingual mixture (-ML). -ML used to be spelled -Multi, which read as multi-vector when it meant multilingual.

NanoVDR-Q-BERT-Qwen3VL2B-2048-ML is a 116M-parameter multilingual text-only query encoder for visual document retrieval. It retrieves document page images as effectively as Vision-Language Models 30-100x its size, with strong cross-lingual transfer across 6 languages.

Built on NanoVDR-Q-DistilBERT-Qwen3VL2B-2048 and further trained with multilingual query augmentation (English + German, French, Spanish, Italian, Portuguese), it is the recommended model for production use with multilingual or mixed-language queries.

Results

Per-Language Teacher Retention

All 6 languages achieve >92% of the 2B teacher's performance.

How It Works

NanoVDR decouples query encoding from document encoding in visual document retrieval:

• Offline indexing: The VLM teacher (Qwen3-VL-Embedding-2B) encodes document page images into single-vector embeddings. This is a one-time cost.
• Online querying: NanoVDR-Q-BERT-Qwen3VL2B-2048-ML encodes text queries in any supported language into the same embedding space via a lightweight text encoder + MLP projector. No vision model needed at query time.

Retrieval uses standard cosine similarity between query and document embeddings.

Usage

from sentence_transformers import SentenceTransformer

# Load the multilingual query encoder
model = SentenceTransformer("nanovdr/NanoVDR-Q-BERT-Qwen3VL2B-2048-ML")

# Encode queries in any supported language
queries = [
    "What was the revenue growth in Q3 2024?",           # English
    "Quel est le chiffre d'affaires du trimestre?",       # French
    "Wie hoch war das Umsatzwachstum im dritten Quartal?", # German
    "Cual fue el crecimiento de ingresos en el Q3?",      # Spanish
]
query_embeddings = model.encode(queries)
print(query_embeddings.shape)  # (4, 2048)

# Retrieve against pre-indexed document embeddings from the VLM teacher
# scores = query_embeddings @ doc_embeddings.T

Full Retrieval Pipeline

from sentence_transformers import SentenceTransformer

# Step 1: Index documents with the VLM teacher (one-time, offline)
from transformers import AutoModel
teacher = AutoModel.from_pretrained("Qwen/Qwen3-VL-Embedding-2B")
# doc_embeddings = teacher.encode(document_images)  # See Qwen3-VL-Embedding docs

# Step 2: Query with NanoVDR-Q-DistilBERT-Qwen3VL2B-2048-ML (online, fast, CPU-only)
student = SentenceTransformer("nanovdr/NanoVDR-Q-BERT-Qwen3VL2B-2048-ML")
query_emb = student.encode("Quel est le chiffre d'affaires?")

# Step 3: Retrieve
scores = query_emb @ doc_embeddings.T
top_k = scores.argsort()[-5:][::-1]

Training Details

• Architecture: google-bert/bert-base-uncased + 2-layer MLP projector (768 → 768 → 2048)
• Training objective: Pointwise cosine alignment with teacher query embeddings
• Training data: 1.49M query-document pairs — 711K original (4 public sources) + 778K machine-translated queries in 5 languages (DE, FR, ES, IT, PT) via Helsinki-NLP Opus-MT models
• Training cost: ~15 GPU-hours on a single H200
• Epochs: 10, lr=3e-4, batch size 1024 (effective)

Multilingual Augmentation Pipeline

1. Extract 489K English queries from training data
2. Translate to 5 target languages using Helsinki-NLP Opus-MT models (~200K per language)
3. Re-encode translated queries with the frozen teacher in text mode to produce target embeddings
4. Combine with original 711K pairs → 1.49M total training samples

Want a system with no teacher at all?

Every model in this table targets Qwen3-VL-Embedding-2B at 2048 dimensions, so documents still have to be indexed by that 2B teacher. A second family targets the 8B teacher at 4096 dimensions and includes a document tower, so nothing multi-billion runs at indexing time either:

The two families are not interchangeable: 2048-d and 4096-d vectors live in different spaces and will not score against each other.

Key Properties

• Output dimension: 2048 (aligned with Qwen3-VL-Embedding-2B)
• Max sequence length: 512 tokens
• Supported languages: English, German, French, Spanish, Italian, Portuguese
• Similarity function: Cosine similarity
• Pooling: Mean pooling
• Normalization: L2-normalized output

Efficiency

Related Models

• NanoVDR-Q-DistilBERT-Qwen3VL2B-2048 — English-focused, same architecture
• NanoVDR-Q-BERT-Qwen3VL2B-2048 — BERT-base backbone (116M)
• NanoVDR-Q-ModernBERT-Qwen3VL2B-2048 — ModernBERT backbone (155M)

Citation

@article{nanovdr2026,
  title={NanoVDR: Distilling a 2B Vision-Language Retriever into a 70M Text-Only Encoder for Visual Document Retrieval},
  author={Liu, Zhuchenyang and Zhang, Yao and Xiao, Yu},
  journal={arXiv preprint arXiv:2603.12824},
  year={2026}
}

License

Apache 2.0