Hugging Face has released LightOn-rerank, a 2B parameter model capable of reranking both text passages and document pages. The model uses a listwise approach to improve performance over traditional pointwise methods. It achieves a NDCG score of 62.66 on the ViDoRe V3 benchmark, surpassing the ColQwen2.5 first stage by 7 nDCG points and leading all other open 2B-class multimodal rerankers measured. The model also remains competitive on text-only BEIR benchmarks.
The development of LightOn-rerank faced several challenges, including the inefficiency of standard speedup techniques used in text reranking. These methods, which split listwise comparisons into smaller subsets, failed to improve performance because the model relies on comparing candidates against each other rather than scoring them independently. Instead, the team found that reducing the number of candidates to rerank significantly improved efficiency, retaining 85% of the rerank lift with 5× fewer windows. Reranking the top-20 candidates instead of 100 achieved this, with top-10 candidates maintaining 58% of the lift at 12× fewer windows.
The model was trained on a mixed batch of text and page inputs, with the base model being Qwen/Qwen3.5-2B, which handles both modalities through a shared decoder and a ViT image encoder. The team observed that while pointwise performance plateaus with model size, listwise ranking continues to scale. At 4B parameters, the model reaches 64.69 NDCG on ViDoRe V3, outperforming the Qwen3-VL-Reranker-8B model at half the parameters.
Source: huggingface
