RenderRank-2B
by nlpai-lab
A 2B reranker that reads documents as rendered images, matching larger text rerankers with about half the input tokens on long documents
nlpai-lab/RenderRank-2BDeploy RenderRank-2B
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
RenderRank-2B is a document reranker from nlpai-lab, fine-tuned from Qwen3-VL-Reranker-2B and released under Apache-2.0 in September 2026. It renders a document's text as page images and scores those against a text query, which takes fewer tokens than the text itself, so more of a long document fits in the same context.
On the card's BEIR reranking test it averages 55.96 NDCG@10 using 290 input tokens per pair, ahead of Qwen3-Reranker-0.6B (54.56 at 439 tokens). On four long-document sets it averages 88.27, level with a 4B reranker while reading about half the tokens.
Architecture
Qwen3-VL-Reranker-2B fine-tuned on rendered text. Documents are drawn as black-on-white pages (Roboto 12 pt at 96 DPI, 896 px wide, up to 56 lines), passed as visual tokens with a text query, and scored as a yes/no relevance probability. It also accepts plain text or existing page images.
Mixpeek SDK Integration
# A reranker scores candidates; index and retrieve the documents with Mixpeek first.
import requests
docs = requests.post(
"https://api.mixpeek.com/v1/retrievers/ret_your_retriever/execute",
headers={"Authorization": "Bearer API_KEY", "X-Namespace": "ns_your_namespace"},
json={"inputs": {"query": "termination notice period"}},
).json()["documents"]Capabilities
- Scores a document as plain text, as rendered page images, or from your own page images
- Multi-page documents scored as one document
- 32,768-token context including query, template and visual tokens
- Yes/no relevance scoring, Apache-2.0
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| BEIR, 11 sets (top 100 BM25 reranked) | NDCG@10 avg | 55.96 | Model card: nlpai-lab/RenderRank-2B (self-reported, 290 input tokens per pair; Qwen3-Reranker-0.6B 54.56 at 439; Qwen3-Reranker-4B 57.99) |
| Long documents (MLDR, 2WikiMQA, QMSum, SummScreenFD) | NDCG@10 avg | 88.27 | Model card (self-reported; Qwen3-Reranker-4B 88.15 using about twice the input tokens) |
| MLDR (English) | NDCG@10 | 99.74 | Model card (4,198 input tokens per pair against 8,863 for Qwen3-Reranker-0.6B) |
Performance
The saving is in input length: rendered documents use roughly a third fewer tokens on BEIR and about half on long documents than the text rerankers in the card's tables. Evaluated in English only. We have not measured it.
Common Pipeline Companions
Frequently Asked Questions
How does RenderRank reduce reranking cost?
It renders document text as page images, which the vision encoder represents with fewer tokens than the original text. On the card's long-document sets it reads about half the tokens of comparable text rerankers at the same accuracy.
Can RenderRank rerank scanned PDFs?
Yes. It takes existing page images directly, one or several pages per document, as well as plain text and text it renders itself.
Is RenderRank free to use commercially?
Yes. It is released under Apache-2.0.
How do I use RenderRank with Mixpeek?
Run a Mixpeek retriever for candidates, then score their text or page images with RenderRank and keep its order, as in the example on this page.
Specification
Research Paper
RenderRank: Learning to Rerank Text with Compressed Visual Tokens
arxiv.orgBuild a pipeline with RenderRank-2B
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free