LiquidAI: LFM2.5-Embedding-350M (free)

LFM2.5-Embedding-350M is LiquidAI’s dense bi-encoder for multilingual and cross-language retrieval. It produces one 1024-dimensional vector for each text and targets short-context search across eleven languages. The named 350M model family uses a bidirectional adaptation of LFM2.5-350M-Base; the detailed model table gives approximately 354 million parameters for this embedding edition.

A bidirectional retrieval representation

The manufacturer adapts the backbone so attention and local convolutions can use context on both sides of a token. CLS-style pooling turns the processed text into one dense vector. Its model table describes seventeen layers: ten convolution blocks, six attention blocks and a pooling stage. This is a retrieval encoder, with its own preparation and output, rather than the text-generation task of the base model.

The separately named LFM2.5-ColBERT-350M retains per-token vectors for late interaction. That representation requires a different index and comparison method. The Embedding model here supplies one vector per document or query, with cosine similarity as the documented comparison.

Use the correct asymmetric prompts

Prepare search queries with “query:” and source passages with “document:”. The manufacturer stores these prompts in its configuration and its reference workflow applies the matching prompt name. It specifically recommends using the corresponding role on each side because the model was trained with these prefixes. This document prefix is different from the “passage:” convention of E5 models.

A practical product-search collection can hold descriptions in the supported languages. Prepare the user’s search wording as a query and each product description as a document. Store the original text, product identifier and language with the vector so a search result can return the corresponding record. For a support collection, retain the question’s relevant policy or instruction instead of replacing it with an unrelated short label.

Languages and passage size

The documented language set is English, Spanish, German, French, Italian, Portuguese, Arabic, Swedish, Norwegian, Japanese and Korean. The manufacturer explicitly targets both retrieval within a language and retrieval across language boundaries. Prepare real questions from the intended audience and inspect the returned document text in those languages.

The document-length specification is 512 tokens. Divide a long source into coherent passages which preserve the heading, definitions and conditions needed by a retrieved result. The role prefix and added context belong to the prepared input. A FAQ answer, a product explanation or a focused section of a technical document is a more suitable unit than a whole oversized manual.

Training and matching

The maker describes three training stages: English contrastive pretraining, multilingual and cross-language distillation, then fine-tuning with hard-mined negatives. The resulting model is intended for product catalogues, FAQ and support collections, file or note search, and enterprise knowledge retrieval. Those are tasks for the text representations, not claims of a particular application’s deployment.

The producer’s example normalises query and document vectors and ranks them with a dot product, equivalent to cosine similarity. Precompute the corpus representations when the source material is prepared, then embed a search query and return candidate documents with their text and identifiers. Keep the selected model and role convention with the index so the numerical comparison uses compatible representations.

The useful result is a compact dense search index or a query vector that retrieves source material across the documented language set. This exact LFM2.5 Embedding edition remains separate from its ColBERT companion, older LFM2 retrievers and externally quantised implementations. The selected passage text is the evidence used by a reader or a later RAG answer stage.