Now loading...
Liquid AI has announced the launch of two innovative multilingual retrieval models, known as LFM2.5-ColBERT-350M and LFM2.5-Embedding-350M. Each model consists of 350 million parameters and is designed to provide efficient multilingual and cross-lingual search capabilities across 11 languages. These models are variations of the previously released LFM2.5-350M-Base, which debuted in March.
The new models are particularly adept at handling short-context searches, such as those found in product catalogs, FAQ knowledge bases, and support documents, making them valuable tools for businesses seeking fast and dependable multilingual searches.
Each model caters to different requirements. The LFM2.5-Embedding-350M model generates a single vector for each document, making it an ideal choice for those prioritizing rapid search capabilities and economical yet valuable indexing. On the other hand, the LFM2.5-ColBERT-350M model translates each individual token into a vector, allowing for more precise matching of queries on a word-by-word basis while also requiring a larger index. This option is recommended when accuracy is of utmost importance.
Both models are derived from the LFM2.5-350M-Base architecture, which has been enhanced with bidirectional components. This adaptation shifts the model’s design from a unidirectional to a bidirectional approach, enabling it to use information from both preceding and succeeding tokens, thus improving retrieval. The architecture retains the efficiency of the original LFM2 design while allowing for comprehensive context representation.
The two models diverge mainly in their text representation methods. The LFM2.5-Embedding-350M relies on CLS-style pooling to yield a single dense embedding, while LFM2.5-ColBERT-350M maintains compact per-token embeddings, allowing for more sophisticated late-stage interactions during the retrieval process.
In terms of training, both models underwent a rigorous three-phase process: large-scale contrastive pretraining in English, multilingual and cross-lingual distillation across all supported languages, and final fine-tuning using challenging negative samples. The training used curated internal data alongside public English retrieval datasets, as well as LLM-based translations to improve multilingual capabilities.
The performance of these models has been evaluated across all 11 supported languages—Arabic, German, English, Spanish, French, Italian, Japanese, Korean, Norwegian, Portuguese, and Swedish—demonstrating leading performance in both multilingual retrieval and cross-lingual question answering tasks. Results suggest both models maintain high retrieval quality consistently across these languages, underlining their adaptability beyond just English.
For deployment, Liquid AI has optimized these models to operate on a range of platforms, releasing modified versions tailored for LFM2.5-ColBERT-350M-GGUF and LFM2.5-Embedding-350M-GGUF. These versions are designed to run efficiently on various devices, from edge devices to laptops, at minimal cost while providing excellent latency. For enterprise-level implementations, Liquid AI has developed an internal GPU infrastructure to ensure low-latency performance under heavy load conditions.
While both models offer powerful performance out of the box, Liquid AI advocates for fine-tuning these models with custom data for domain-specific retrieval applications. This guidance is especially emphasized for LFM2.5-Embedding-350M, where users can find simplistically laid out fine-tuning examples in the Hugging Face model card. The new LFM2.5 retrieval models are now accessible on Hugging Face, and organizations looking to leverage these tools for extensive retrieval solutions are encouraged to get in touch to learn more.
