MongoDB - Build AI That Scales
Mini toolsLFM2.5-ColBERT-350M
By Liquid AI
Model family: LFM
LFM2.5-ColBERT-350M is a bidirectional encoder adapted from the LFM2.5-350M-Base causal decoder. It replaces causal attention masks with bidirectional ones and makes short convolutions non-causal, enabling full-context token representations. Each token is converted to a compact vector for MaxSim late interaction scoring, trading a larger index for higher retrieval accuracy and better generalization over single-vector models. Trained in three stages: large-scale English contrastive pretraining, multilingual and cross-lingual distillation from a strong teacher model, and hard-negative fine-tuning. Supports Arabic, German, English, Spanish, French, Italian, Japanese, Korean, Norwegian, Portuguese, and Swedish. Achieves best-in-class results on NanoBEIR Multilingual and MKQA-11. Suited for product catalogs, FAQ bases, and support docs. GGUF available for llama.cpp on CPUs, laptops, and edge devices.
New
Text
Gen 7
Overview
350M-parameter bidirectional multilingual retrieval model using per-token ColBERT-style late interaction (MaxSim) for accurate multilingual and cross-lingual search. Supports 11 languages. Built from LFM2.5-350M-Base with bidirectional attention and non-causal convolutions. Available in GGUF format for CPU and edge deployment.
About Liquid AI
Liquid AI is an MIT spin-off building efficient general-purpose AI models (Liquid Foundation Models, or LFMs) that run on edge devices with less memory and power.
They recently raised $250M in Series A funding to scale model development and deployment.
Tools using LFM2.5-ColBERT-350M
No tools found for this model yet.
Last updated: June 19, 2026
