LightOn-rerank models are unified cross-encoder rerankers: Qwen3.5 LoRAs scoring both text passages and document page images against a query.