📝 Blog post

LightOn-rerank-PW-4B

Unified Text + Visual Document Reranker by LightOn

PW-0.8B | LW-0.8B | PW-2B | LW-2B | PW-4B | LW-4B

---

About the LightOn-rerank family

Production retrieval pipelines usually need two rerankers: one for text passages and one for visual documents (PDF pages, slides, scans). LightOn-rerank models are unified cross-encoder rerankers: a single model scores both text passages and document page images against a query, on top of any first-stage retriever (BM25, dense embeddings, or ColPali-family late-interaction models).

The models are built on Qwen3.5 backbone (hybrid linear + full attention) and jointly fine-tuned on text and visual reranking data with mixed-modality batches (LoRA, merged into the released weights). Training data is English-only; French performance transfers zero-shot from the multilingual backbone.

The family comes in two scoring flavours × three sizes (0.8B / 2B / 4B):

• PW (pointwise): each candidate is scored independently. The model judges whether the document answers the query, and the score is logit("Yes") − logit("No"). One forward pass per candidate and no generation.
• LW (listwise): generative listwise ranking, where 4 candidates are placed in a single prompt and the model generates a permutation ([2] > [4] > [1] > [3]). Larger candidate pools are ranked with a sliding window (window 4, stride 2, bottom-to-top). Cross-document attention makes LW markedly stronger on hard visual reranking, and unlike pointwise scoring it keeps improving with backbone size.

LightOn-rerank-PW-4B is the 4B pointwise member of the family. On visual reranking it ties the 2B pointwise model (59.80 vs 59.87 nDCG@10 on ViDoRe V3): direct evidence that independent Yes/No scoring is a capacity bottleneck that extra parameters cannot fix. On text BEIR, however, it posts the strongest scores in the family (50.19 decontaminated mean). For 4B-budget vision-heavy deployments, LightOn-rerank-LW-4B is +4.9 nDCG@10 at the same size.

Results

ViDoRe V3 (visual document reranking, 8 domains × EN/FR queries), overall nDCG@10, ColQwen2.5-v0.2 first stage, retrieve 100 / rerank 100. All models, including baselines, were re-evaluated under this same two-stage protocol, so numbers are mutually comparable but not comparable to vendor-reported end-to-end results.

Model	Params	Scoring	ViDoRe V3 overall nDCG@10
LightOn-rerank-LW-4B	4.5B	listwise	64.69
Qwen3-VL-Reranker-8B	8B	pointwise (pooling)	64.23
LightOn-rerank-LW-2B	2.2B	listwise	62.66
LightOn-rerank-PW-2B	2.2B	pointwise	59.87
LightOn-rerank-PW-4B (this model)	4.5B	pointwise	59.80
jina-reranker-m0	2.4B	pointwise	59.40
Qwen3-VL-Reranker-2B	2B	pointwise (pooling)	59.18
LightOn-rerank-LW-0.8B	0.85B	listwise	58.25
MonoQwen2-VL-v0.1	2B	pointwise	57.76
First-stage only (ColQwen2.5, no rerank)	—	—	55.60
LightOn-rerank-PW-0.8B	0.85B	pointwise	48.20

ViDoRe V3 detail (nDCG@10, ColQwen2.5 first stage, rerank-100)

Domain	EN	FR
finance_en	70.98	68.54
finance_fr	45.96	37.92
computer_science	79.25	74.71
hr	58.67	53.54
energy	70.95	72.36
industrial	60.54	47.33
pharmaceuticals	68.11	67.70
physics	44.46	35.77
mean	62.37	57.23

Overall nDCG@10: 59.80 (EN 62.37 / FR 57.23).

BEIR results (text reranking)

13 datasets, nDCG@10, BM25 first stage, retrieve 100 / rerank 100, same protocol as all other family members. ⚠️ marks datasets in the text training mix (NQ, MSMARCO); the decontaminated mean excludes them.

Dataset	nDCG@10
fever	81.77
scifact	76.13
trec-covid	79.38
hotpotqa	70.79
nq ⚠️	61.03
dbpedia	44.21
arguana	41.47
fiqa	40.91
msmarco ⚠️	39.20
nfcorpus	37.41
touche-2020	35.68
climate-fever	26.46
scidocs	17.87
Mean (13)	50.18
Decontaminated mean (11, excl. ⚠️)	50.19

This is the best text result in the LightOn-rerank family (2B pointwise: 49.13, 4B listwise: 48.33 decontaminated mean). Pointwise plateaus with scale on vision but not on text. If your workload is text-dominant and you have the 4B budget, this is the family's strongest text reranker.

Model Details

• Model type: multimodal cross-encoder reranker (pointwise: each candidate is scored independently as logit("Yes") − logit("No"))
• Base model: Qwen/Qwen3.5-4B (Qwen3.5 hybrid linear + full attention VLM)
• Parameters: ≈4.5B (bfloat16, 9.1 GB)
• Inputs: query (text) + candidate document(s): text passage or page image
• Fine-tuning: joint text+vision LoRA (r=32, α=32, rsLoRA, merged into the released weights), mixed-modality batches (2 text + 2 vision groups per micro-batch), vision loss weight 1.3, lr 1e-4, 1 epoch (465 steps)
• Data: same 213k groups as the listwise models — 107k text groups (NQ, TriviaQA, MS MARCO; each a [pos, neg_0, neg_1, neg_2] 4-list with hard negatives mined via the NV-Retriever approach with GTE-ModernBERT) + 106k vision groups (ColPali train set with negatives mined by Nomic). For pointwise training each group is flattened into (query, document, Yes/No) triples.
• Languages: English (training), French (zero-shot transfer)
• Requirements: transformers >= 5.4.0 (qwen3_5 architecture); sentence-transformers >= 5.4.0 for the CrossEncoder usage

Usage: pointwise reranking

Each candidate is scored independently as logit("Yes") − logit("No") at the first generated position; sort candidates by descending score. The model was trained with a fixed system prompt and user template: use them verbatim for best results.

Using Sentence Transformers

Install Sentence Transformers (>= 5.4.0):

pip install "sentence-transformers[image]"

The trained system prompt and user templates are baked into the bundled reranker chat template — including the non-thinking generation prompt, so no enable_thinking handling is needed — and query-document pairs are formatted correctly out of the box:

from sentence_transformers import CrossEncoder

model = CrossEncoder("lightonai/LightOn-rerank-PW-4B")

query = "What is late interaction in neural information retrieval?"
documents = [
    "ColBERT computes token-level query-document interactions at search time...",
    "The Eiffel Tower is located on the Champ de Mars in Paris.",
]

pairs = [(query, doc) for doc in documents]
scores = model.predict(pairs)
print(scores)
# [-3.     -8.0625]

rankings = model.rank(query, documents)
print(rankings)
# [{'corpus_id': 0, 'score': -3.0}, {'corpus_id': 1, 'score': -8.0625}]

To rerank page images, pass a PIL.Image (or an image URL or file path string) as the document. Text and image candidates can be mixed in the same call:

from PIL import Image

pairs = [
    (query, Image.open("page_1.png")),
    (query, "https://example.com/page_2.png"),
    (query, "A text passage candidate for the same query."),
]
scores = model.predict(pairs)

Scores are raw logit("Yes") − logit("No") differences. You can map them to 0...1 probabilities with model.predict(pairs, activation_fn=torch.nn.Sigmoid()).

Using Transformers

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "lightonai/LightOn-rerank-PW-4B"
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",  # optional, remove if flash-attn is not installed
    device_map="cuda",
).eval()
processor = AutoProcessor.from_pretrained(model_id)
processor.tokenizer.padding_side = "left"  # scores are read at the last position

YES_TOKEN_ID = 9175  # "Yes"
NO_TOKEN_ID = 2665   # "No"

SYSTEM_PROMPT = "Judge whether the document is relevant to the query. Answer Yes or No."
USER_TEMPLATE = (
    "Given a query, determine if the document is relevant. "
    "The query is: {query}\n\nDocument: {doc}"
)

query = "What is late interaction in neural information retrieval?"
documents = [
    "ColBERT computes token-level query-document interactions at search time...",
    "The Eiffel Tower is located on the Champ de Mars in Paris.",
]

texts = [
    processor.apply_chat_template(
        [
            {"role": "system", "content": SYSTEM_PROMPT},
            {"role": "user", "content": USER_TEMPLATE.format(query=query, doc=doc)},
        ],
        tokenize=False,
        add_generation_prompt=True,
        enable_thinking=False,  # REQUIRED on the 4B backbone (thinking on by default)
    )
    for doc in documents
]
inputs = processor(
    text=texts, return_tensors="pt", padding=True, truncation=True, max_length=2048
).to(model.device)
with torch.inference_mode():
    logits = model(**inputs).logits[:, -1]
scores = (logits[:, YES_TOKEN_ID] - logits[:, NO_TOKEN_ID]).tolist()
# [-3.0, -8.0625]

ranked = sorted(zip(scores, documents), reverse=True)

To score a page image instead of a text passage, replace the user message with:

VISION_TEMPLATE = "Given a query, determine if the document image is relevant. The query is: {query}"

{"role": "user", "content": [
    {"type": "image", "image": page_image},  # PIL.Image
    {"type": "text", "text": VISION_TEMPLATE.format(query=query)},
]}

and pass images=[page_image, ...] to the processor call (keep the same system prompt).

Serving with vLLM

vllm serve lightonai/LightOn-rerank-PW-4B --trust-remote-code --max-model-len 16384

resp = client.chat.completions.create(
    model="lightonai/LightOn-rerank-PW-4B",
    messages=messages,          # same system + user messages as above
    max_tokens=1,
    logprobs=True,
    top_logprobs=20,
    temperature=0.0,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},  # REQUIRED
)
top = resp.choices[0].logprobs.content[0].top_logprobs
lp = {t.token: t.logprob for t in top}
score = lp.get("Yes", -100.0) - lp.get("No", -100.0)

Full-page document images can exceed 8k tokens, so keep --max-model-len at 16384 or higher when reranking page images.

Notes & limitations

• Thinking must be disabled for scoring. Qwen3.5-4B's chat template enables <think> by default; with it on, the first generated token is a thinking token and the Yes/No logprobs are distorted. Pass enable_thinking=False to apply_chat_template (or chat_template_kwargs={"enable_thinking": False} via the vLLM OpenAI client) as shown above. The Sentence Transformers path handles this automatically: the bundled reranker chat template hardcodes the non-thinking generation prompt.
• Pointwise scoring does not benefit from the 2B→4B scale-up on vision (59.80 vs 59.87 for the 2B), while generative listwise gains +2.0 points over the same step. For vision workloads prefer LightOn-rerank-LW-4B at this size, or LightOn-rerank-PW-2B for the same visual quality at lower cost.
• Training data is English-only. French works zero-shot (the backbone is multilingual) but is slightly behind English on average.
• BEIR contamination flag: NQ and MSMARCO are part of the text training data; headline text figures use decontaminated means that exclude them.

The LightOn-rerank family

Model	Backbone	Scoring	ViDoRe V3 overall nDCG@10
LightOn-rerank-PW-0.8B	Qwen3.5-0.8B	pointwise	48.20
LightOn-rerank-LW-0.8B	Qwen3.5-0.8B	listwise	58.25
LightOn-rerank-PW-2B	Qwen3.5-2B	pointwise	59.87
LightOn-rerank-LW-2B	Qwen3.5-2B	listwise	62.66
LightOn-rerank-PW-4B	Qwen3.5-4B	pointwise	59.80
LightOn-rerank-LW-4B	Qwen3.5-4B	listwise	64.69

Rule of thumb: LW models are stronger at every size (and the gap grows with size); PW models are cheaper to serve and score candidates independently. For the best quality pick LW-4B; for the best quality/cost trade-off pick LW-2B; for maximum throughput on text-heavy workloads pick a PW model.

Citation

@misc{ananya2026lightonrerank,
  title={One Adapter, Both Modalities: Field Notes from Building and Serving a Multimodal Reranker},
  author={Ananya, Ishrat Jahan and Chatelain, Amelie},
  year={2026},
  howpublished={\url{https://huggingface.co/blog/lightonai/lighton-rerank}},
}