Liquid AI releases LFM2.5 retrieval models for fast multilingual search

Liquid AI released two new retrieval models on Hugging Face: LFM2.5-ColBERT-350M, a late-interaction model, and LFM2.5-Embedding-350M, a dense bi-encoder. Both carry 350M parameters and are designed for fast multilingual and cross-lingual search across 11 languages, with a small footprint that suits a range of deployment scenarios from edge to server. They are the first bidirectional members of the LFM family, building on the earlier LFM2.5-350M-Base, and ship under the LFM Open License v1.0.

The pairing of a dense bi-encoder and a ColBERT-style late-interaction model is a deliberate offering of two retrieval paradigms: bi-encoders are fast and cheap for first-stage retrieval, while late-interaction models like ColBERT trade some efficiency for higher accuracy by preserving token-level interactions. At just 350M parameters, both are aimed at teams that want strong multilingual retrieval without the cost of large embedding models.

The release lands amid a vibrant week for open models on Hugging Face: Gemma 4 12B became the platform's most-downloaded model with 4M+ downloads in its first week, GLM 5.2 generated buzz (Hugging Face's Clement Delangue amplified a claim it's 'a real frontier' model running locally against Claude Opus), and Transformers v5.12.0 added the 428B-parameter MiniMax-M3-VL. The retrieval-model niche is competitive, with strong incumbents from Cohere, BAAI (BGE), and Nomic.

For practitioners building RAG systems, small efficient multilingual retrievers are valuable precisely as the industry pushes agentic, retrieval-grounded applications (see AWS's AgentCore web search and managed knowledge bases this week). The open license and small size lower the barrier for self-hosted, privacy-sensitive deployments. Watch for independent benchmark comparisons against BGE-M3 and Cohere's multilingual embeddings to gauge real-world quality.