This week Liquid AI released two 350‑million‑parameter retrieval models, LFM2.5-ColBERT-350M and LFM2.5-Embedding-350M. They are the first bidirectional members of the LFM family and work across eleven languages. Teams that need fast multilingual search often face three problems: limited hardware for large indexes, a trade‑off between search speed and relevance quality, and the effort to swap models into an existing RAG pipeline without rewriting code. The Embedding model solves the first problem by turning each document into a single 1024‑dimension vector, giving the smallest possible index and sub‑10 ms query latency on a laptop when embeddings are cached. The ColBERT model keeps a 128‑dimension vector per token, which lets it match queries word‑by‑word for higher accuracy while still running under 35 ms on the same hardware if document vectors are computed on the fly. Both models ship as GGUF files for llama.cpp, so they run on CPUs, laptops and edge devices, and they also scale to sub‑millisecond latency on an H100 GPU stack. Integration is straightforward: the Embedding model works with sentence‑transformers using the query: and document: prompts, and the ColBERT model works with PyLate’s PLAID index or can rerank a first‑stage list without building an index. Because they are released under the LFM Open License v1.0, teams can drop them into current pipelines, fine‑tune on domain data, and deploy anywhere from on‑device search to enterprise knowledge assistants. The result is multilingual, cross‑lingual retrieval that is fast, accurate and cheap to operate. #AI #Product #Search #LLM #RAG #EdgeComputing