Liquid AI has just released two multilingual retrieval models that could shift the balance for anyone building semantic search on proprietary data. LFM2.5-Embedding-350M and LFM2.5-ColBERT-350M are designed as drop-in replacements for existing RAG pipelines, with a focus on computational efficiency that makes them especially suited to on-premise deployments.

Two approaches, one goal: multilingual accuracy

The first, LFM2.5-Embedding-350M, is a dense bi-encoder that outputs a single vector per document. The resulting index is as compact as possible, yielding lightning-fast searches with minimal resource consumption. The second, LFM2.5-ColBERT-350M, takes a late-interaction approach: it stores one vector per token and uses MaxSim to match queries against documents. This enables a compelling scenario: you can store documents in one language (say, product descriptions in English) and retrieve them accurately with queries in many other languages. Both are optimized for eleven languages and, according to benchmarks, deliver top-tier accuracy for models of their size.

LFM2.5: efficiency as the enabler

The two retrievers are built on the LFM2 architecture, already known for its inference speed. This efficient backbone lets the models run at speeds comparable to much smaller ones, lowering the hardware barrier. You don’t need enterprise-grade GPUs to achieve quality cross-lingual retrieval: the GGUF version distributed on Hugging Face is ready for consumer hardware or low-VRAM servers – a detail that matters for companies that want to keep data inside their own perimeter.

Data sovereignty through local embeddings

These releases fit a broader trend: the rising demand for AI infrastructures that do not rely on external cloud APIs. In regulated industries or privacy-critical settings, running the entire retrieval pipeline – from embedding to augmented generation – on your own hardware isn’t just a preference; it’s a requirement. Liquid AI’s models, with their reduced footprint and quantized format, directly address this need. It’s no coincidence that the GGUF format was available from day one: built for local execution with tools like llama.cpp, it bridges the gap between academic research and practical production use without cloud intermediaries.

The trade-off to consider

Choosing between the two models isn’t trivial and requires thinking about project goals. LFM2.5-Embedding-350M offers extremely lean indexes and minimal response times but trades off some granularity: a single vector per document may be less precise at capturing nuanced terminology. LFM2.5-ColBERT-350M, by contrast, with one vector per token, raises accuracy at the expense of heavier storage. In an on-premise scenario, this means balancing storage hardware costs (disks and RAM) against search quality. For fine-tuning or adaptation to specialized domains, the compact format of the Liquid models further reduces training time and required resources – a plus for infrastructure teams.

A forward-looking perspective

Liquid AI is no stranger to transparency: releasing open models in a quantized format signals a strategy aimed at winning over developers and enterprises that demand full control over their stack. At a time when Large Language Models dominate the conversation, it’s crucial to remember that retrieval quality is the first brick of a trustworthy RAG system. These two embedding models, thanks to their lightness and native multilingualism, could become a standard component in on-premise architectures for anyone who needs to search across languages without ever letting data leave home.