Liquid AI Introduces LFM2.5-Embedding-350M and LFM2.5-ColBERT-350M: Dense Bi-Encoder and Late-Interaction Models for Fast Multilingual Search Across 11 Languages

Liquid AI has unveiled two new retrieval models designed to enable fast, efficient multilingual search on resource-constrained devices. The LFM2.5-Embedding-350M and LFM2.5-ColBERT-350M represent a significant advancement in making sophisticated search capabilities accessible without requiring extensive computational infrastructure. These models combine distinct architectural approaches to deliver both speed and accuracy across 11 languages, addressing a critical gap in edge computing deployment.

Liquid AI's LFM2.5 Retrievers employ a dual-architecture strategy to optimize search performance. The LFM2.5-Embedding-350M functions as a dense bi-encoder model, converting queries and documents into vector representations for rapid similarity matching. The LFM2.5-ColBERT-350M implements late-interaction technology, which performs token-level matching after encoding to preserve granular semantic information. Both models operate at 350 million parameters, making them lightweight enough for deployment on edge devices while maintaining multilingual support across 11 languages. This combination allows developers to choose between speed and precision based on their specific application requirements.

• Edge deployment capability: Models small enough for on-device inference reduce latency and eliminate cloud dependency for search operations
• Multilingual accessibility: Support for 11 languages enables global applications without requiring separate language-specific models
• Flexibility in search approaches: Offering both dense and late-interaction architectures allows developers to optimize for their particular use cases
• Cost reduction: Lighter models decrease infrastructure requirements and associated operational expenses
• Privacy enhancement: On-device processing minimizes data transmission to external servers

The introduction of these models reflects growing demand for artificial intelligence capabilities that function efficiently beyond data centers. As organizations increasingly prioritize privacy, latency reduction, and cost optimization, retrieval models that perform effectively on edge devices become essential infrastructure. Liquid AI's focus on multilingual support particularly positions these models for global applications, from mobile devices to IoT systems. The dual-architecture approach demonstrates how modern AI engineering balances competing demands between speed and accuracy, enabling developers to make informed choices about their search infrastructure without sacrificing performance or accessibility.