NanoVDR¶

Development line: project:nanovdr · thread nanovdr-public-release
Last event: 2026-03-17 · 1 dated since 2026-03-17 · Researched: 2026-09-05 · confidence: high

What it is¶

NanoVDR is a visual-document retrieval system for teams searching PDF pages, reports, and page images. It acts as a query-side replacement rather than a standalone document indexer.

• Document indexer: indexes pages offline with Qwen3-VL-Embedding-2B.
• Query encoder: embeds text queries with 69–151M-parameter student models.
• Scorer: ranks compatible vectors with a dot product.

The 69M multilingual query model runs at 51 ms per CPU query and requires teacher-indexed documents.

Development line¶

• 2026-03-17 — NanoVDR linked public Hugging Face resources and a demo. On 2026-03-17, the line published a Hugging Face blog article, a project page, and a demo space. This established discoverable project and demonstration resources for the public release, though we did not research their contents for this review.

What changed¶

• 2026-03-17 — NanoVDR was presented as a 69M text-only query encoder distilled from Qwen3-VL-Embedding-2B for visual-document retrieval. The first-party article is dated 2026-03-16 and reports 95.1% teacher retention for NanoVDR-S-Multi.
• 2026-08-11 — Checkpoint names changed to encode tower, teacher, and embedding width. Old links redirect.
• 2026-08-29 — The team posted arXiv v3. The project repository distinguishes the later DistilVDR document-tower work from the original NanoVDR query-tower line.

How to use this¶

Check the linked Hugging Face resources and public demo before evaluation as of 2026-03-17; specific capabilities, requirements, and results remain unverified.

1. Load a compatible query checkpoint with SentenceTransformers and encode the text query. — https://github.com/Ryenhails/NanoVDR
2. For the original NanoVDR line, index pages offline with Qwen3-VL-Embedding-2B, then rank its page vectors with the query vector using a dot product. — https://huggingface.co/nanovdr
3. Use the hosted demo only for exploratory queries; it may be asleep until restarted. — https://huggingface.co/spaces/nanovdr/NanoVDR-Demo

Best practices¶

• Match query and document towers by both teacher and embedding width; vectors from different teacher spaces are not interchangeable. — https://github.com/Ryenhails/NanoVDR
• Prefer the multilingual -ML query checkpoint unless an English-only model is specifically required. — https://github.com/Ryenhails/NanoVDR
• Use the original line for CPU query serving with teacher-built indexes; use the DistilVDR pair only when teacher-free indexing is required. — https://github.com/Ryenhails/NanoVDR

Superseded by this¶

• 2026-08-11 — Checkpoint names stating tower, teacher, and vector width replace earlier names; legacy model links redirect.

Still unknown¶

• The event_findings and new_events fields are absent from the response schema, so the dated additions stay in what_changed.
• A first-party article dated 2026-03-16 corroborates the 2026-03-17 event.
• The 2026-08-11 rename and 2026-08-29 arXiv v3 belong to later updates, not the 2026-03-17 event.
• The public demo was sleeping when checked, so an interactive run is unverified.

Sources¶

Agent brief¶

• Subject: project:nanovdr, thread nanovdr-public-release, 1 dated events 2026-03-17 → 2026-03-17.
• Practical note: As of 2026-03-17, practitioners should consult NanoVDR’s linked Hugging Face project resources and public demo before evaluating it; specific capabilities, requirements, and results remain unverified.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.