Liquid AI released LFM2.5-Retrievers, a family of 350M-parameter bidirectional retrieval models. The standout: state-of-the-art performance on 11-language retrieval benchmarks, with a model small enough to run on a Raspberry Pi.
The technical details: LFM2.5-Retrievers use a bidirectional Transformer architecture (vs the unidirectional LFM-2 series), with 350M parameters. The bidirectional design is critical for retrieval — the model needs to attend to both directions of the query and document to compute the relevance score. The "11-language" training corpus includes English, Chinese, Japanese, Korean, Arabic, Hindi, French, German, Spanish, Portuguese, and Russian.
The benchmark: on the BEIR benchmark (11 languages), LFM2.5-Retrievers-350M hits an average nDCG@10 of 58.7, beating the previous SOTA (BGE-M3-568M) by 2.1 points while being 40% smaller. On the MIRACL benchmark (cross-lingual retrieval), it also hits SOTA, with a 3.5-point improvement.
The "Raspberry Pi deployment" highlight: the 350M model is small enough to run on a Raspberry Pi 5 (8GB RAM) at 50ms per query, with INT8 quantization. This makes on-device multilingual retrieval a real possibility for privacy-sensitive use cases (legal, medical, personal notes).
The bigger takeaway: "specialized small models" are beating "general large models" in retrieval. The 350M LFM2.5-Retriever beats the 568M BGE-M3, which beats the 7B generic embedding models. The "specialization tax" is real — a model trained for a specific task is significantly better than a general model of the same size. For the industry, this means "task-specific embedding models" are the right choice for production deployments, not general-purpose embeddings.