Liquid AI releases LFM2-ColBERT-350M multilingual retriever
Liquid AI's new 350M-parameter late-interaction retriever lets documents stored in one language be searched with queries in others, a common gap for multilingual search systems.
Liquid AI has released LFM2-ColBERT-350M, a late-interaction retriever built on its LFM2 backbone. The model is available on Hugging Face along with an interactive demo.
Late-interaction models encode queries and documents separately at the token level, then compare them at query time, for example with MaxSim. Documents can be embedded ahead of time, as with bi-encoders, but the token-level comparison keeps more of the precision of a cross-encoder. That lets one model handle both retrieval and ranking.
Liquid compared the model with GTE-ModernColBERT-v1 (150M parameters), which it calls the strongest late-interaction retriever under 500M parameters. To do this, it extended the NanoBEIR benchmark to cover Japanese and Korean and published the dataset as LiquidAI/nanobeir-multilingual-extended.
The company says the new model is stronger across languages, especially German, Arabic, Korean and Japanese. Its English score is close to the baseline's: NDCG@10 of 0.661 versus 0.680.
The largest gains show up when the query and the document are in different languages. Some NDCG@10 scores on English documents, LFM2-ColBERT-350M first and GTE-ModernColBERT-v1 second:
- German queries: 0.554 vs 0.408
- Korean queries: 0.395 vs 0.073
- Arabic queries: 0.416 vs 0.042
Liquid says cross-lingual results are best among English, French, Spanish, Italian, Portuguese and German.
LFM2-ColBERT-350M has more than twice as many parameters as GTE-ModernColBERT-v1. Even so, Liquid says it matches that model's speed for encoding both queries and documents across batch sizes. The company credits LFM2's design, which pairs gated short-range convolutions with grouped-query attention. The post did not include raw throughput figures.