LightOn-rerank

Open weights · LightOn · Updated 16 Oct 2026

LightOn-rerank is a family of six rerankers that score text passages and document page images with the same model, so one second stage covers both a text index and a PDF-page index. It comes in three sizes (0.8B, 2B, 4B), each in a pointwise and a listwise version, built on Qwen3.5 and released under Apache 2.0. On visual documents the listwise 4B is the strongest reranker in LightOn’s own comparison; on text it trails jina-reranker-m0.

Pointwise or listwise

Each size comes in two versions, and the choice matters more than the size:

• PW (pointwise) scores each candidate on its own, as the difference between the “Yes” and “No” logits. One forward pass per candidate, easy to batch, and it loads into sentence-transformers’ CrossEncoder.
• LW (listwise) reads four candidates in one prompt and generates their order, such as [2] > [4] > [1] > [3]. Longer lists are ranked with a sliding window of four, moving two at a time from the bottom up. Seeing candidates side by side makes it stronger on visual documents at every size, at roughly three times the latency of pointwise.

Training data is English only. French works without extra training, through the multilingual backbone, but trails English slightly.

Scores, and whose they are

LightOn’s own runs, from the model cards, with every model re-evaluated under the same protocol. ViDoRe V3: NDCG@10 over the top 100 page images from ColQwen2.5. BEIR: NDCG@10 over BM25’s top 100, averaged over the 11 datasets that weren’t in LightOn’s training data. Neither matches vendors’ own numbers: Qwen’s card, for instance, gives the 8B a higher ViDoRe V3 score from its own setup (see the Qwen3-VL-Reranker page).

Three things the table shows. The listwise models gain with size, and LW-4B edges past Qwen3-VL-Reranker-8B at about half the parameters; the pointwise 4B does no better than the pointwise 2B. On text, the three LightOn models with published scores sit three to four points behind jina-reranker-m0, whose weights are non-commercial. And PW-0.8B scores below the first stage it reranks, so don’t use it on page images.

Quick start

Pointwise, with sentence-transformers

pip install "sentence-transformers[image]"   # 5.4 or later

from sentence_transformers import CrossEncoder

model = CrossEncoder("lightonai/LightOn-rerank-PW-2B")

query = "What is late interaction in neural information retrieval?"
documents = [
    "ColBERT computes token-level query-document interactions at search time...",
    "https://example.com/page_2.png",   # a page image: URL, path or PIL.Image
]
rankings = model.rank(query, documents)

Text and image candidates can share a call. The chat template the model was trained with ships inside the checkpoint, so pairs are formatted correctly without extra code.

Listwise models

The LW models don’t load into CrossEncoder: they generate an order rather than a score. The model cards include a short transformers implementation of the four-candidate window and the sliding pass over a longer list. To serve either kind with vLLM, turn thinking off (enable_thinking: False; the Qwen3.5 template has it on by default and it distorts the scores), and allow at least 16,384 tokens of context when ranking full page images.

When to pick LightOn-rerank

• Yes: you search PDF pages, slides or scans, and want a commercial-friendly licence: LW-2B for the best quality for its cost, LW-4B for the best quality.
• Yes: one model has to rerank both a text index and a page-image index.
• Maybe not: text only. Text rerankers such as Qwen3-Reranker are the more natural fit there.
• Maybe not: languages other than English and French, which it wasn’t trained or tested on.

Pros and cons

Other models