Lychee-Reranker
An open reranker family built on the Fast but Not Late Interaction design, released as KaLM-Reranker-V1 and renamed Lychee-Reranker in later revisions of its paper (Lychee-Reranker - Fast but Not Late Interaction for Compressed Document Reranking). Models are on Hugging Face under the KaLM-Embedding collection; code is at https://github.com/HITsz-TMG/Lychee-Reranker.
- Model card (Large R2): https://huggingface.co/KaLM-Embedding/KaLM-Reranker-V1-Large-R2
- Paper: https://arxiv.org/abs/2606.22807
Models
Three sizes on T5Gemma2 encoder-decoders, by activated parameters: Nano (0.27B), Small (1B) and Large (4B). The R2 models keep the same architecture and inference interface as the original release.
Usage
The document is wrapped as <Document>: {document} and fed to the encoder, where it is compressed by mean-pooling groups of encoder token states (the model card’s Sentence Transformers example uses chunk_size=4). The decoder receives the instruction and query as a chat-style prompt and answers “yes” or “no”. The score is P(yes) by default, or yes_logit minus no_logit with an Identity activation.
Performance
On BEIR, Large averages 63.53 nDCG@10 (Qwen3-Reranker-4B: 63.50, Qwen3-Reranker-8B: 65.11); on MIRACL Large averages 74.92. Estimated online cost is a small fraction of the Qwen3-Reranker series at comparable quality (Small 4.6x vs 28.2x for Qwen3-Reranker-0.6B). Details and caveats in the paper note.
Related Concepts
- Fast but Not Late Interaction · Reranking · Cross-Encoder · Late Interaction · Matryoshka Embeddings