Liquid AI releases two multilingual retrieval models

Liquid AI announced two retrieval models, LFM2.5-Embedding-350M and LFM2.5-ColBERT-350M. Both are built on LFM2.5-350M-Base, a checkpoint released in March, and target short-context search such as product catalogs, FAQ knowledge bases and support docs.

The two models trade off index size against accuracy. The Embedding model turns each document into a single vector, which gives the fastest search and the smallest index. The ColBERT model keeps one vector per token and matches queries word by word using MaxSim late interaction. Liquid says this is more accurate but needs a larger index.

To turn the causal decoder into an encoder, Liquid replaced the causal attention mask with a bidirectional one and made the LFM2 short convolutions non-causal. Training ran in three stages: contrastive pretraining in English, multilingual and cross-lingual distillation from a teacher model, and fine-tuning on hard-mined negatives. Liquid used LLM-based translation to create more multilingual training pairs and credited LightOn's LateOn and DenseOn release as inspiration for the staged recipe.

The supported languages are Arabic, German, English, Spanish, French, Italian, Japanese, Korean, Norwegian, Portuguese and Swedish. On the company's NanoBEIR Multilingual Extended results (NDCG@10), the ColBERT model averaged 0.605 and the Embedding model 0.577, compared with 0.556 for Qwen3-Embedding-0.6B and 0.540 for Liquid's earlier LFM2-ColBERT-350M. On MKQA-11 cross-lingual QA (Recall@20), the two models scored 0.694 and 0.691. The company did not publish full comparison figures for that benchmark. All benchmark results are self-reported.