Liquid AI drops lightweight multilingual search models

Liquid AI just released LFM2.5-ColBERT-350M and LFM2.5-Embedding-350M, two 350M-parameter retrieval models built for low-cost, high-speed search across 11 languages. They are small enough to run on laptops and edge devices, making them ideal for short-context tasks like product catalogs, customer support docs, and FAQ repositories.

What the models do

Both models cover Arabic, German, English, Spanish, French, Italian, Japanese, Korean, Norwegian, Portuguese, and Swedish, but they serve two distinct retrieval strategies.

LFM2.5-Embedding-350M generates a single dense vector for each document using pooling. It delivers the fastest query throughput and the smallest storage footprint. If you need cheap indexing and blazing-fast response times on massive collections, this is the bi-encoder to pick.

LFM2.5-ColBERT-350M takes the late-interaction route, assigning compact vectors to individual tokens. It performs token-by-token comparisons at query time. That eats up more storage for your index, but the payoff is significantly better retrieval precision and stronger generalization across language boundaries.

Turning a causal backbone bidirectional

Retrieval models require bidirectional context, but standard causal autoregressive models only look from left to right. Liquid AI adapted its existing LFM2.5-350M-Base checkpoint from March by applying bidirectional architectural patches.

The team replaced the causal attention mask with a bidirectional mask, allowing every token to look both backward and forward. They also altered the architecture's short convolutions to be non-causal, letting them gather local representations symmetrically.

After rebuilding the backbone as an encoder, both models underwent a three-stage training recipe: large-scale contrastive pretraining on English data, cross-lingual distillation across all 11 target languages, and hard-negative mining fine-tuning.

Benchmarks and running it locally

Despite their small size, both models punched well above their weight class on standardized evaluations. On the NanoBEIR multilingual benchmark, LFM2.5-ColBERT scored 0.605 and the Embedding model hit 0.577, outperforming larger alternatives like the 600M Qwen3-Embedding-0.6B at 0.556 and Alibaba's gte-multilingual-base at 0.528.

Cross-lingual performance held up equally well. On MKQA-11, the ColBERT model posted a 0.694 Recall@20 score while the Embedding model posted 0.691, leading all tested baselines across the board.

Deployment is straightforward. Weights are live on Hugging Face, alongside quantized GGUF builds for llama.cpp. On regular laptop CPUs, single-query latency runs around 7 to 8 milliseconds when documents are cached. On enterprise GPUs, query embedding takes roughly 1.3 to 1.5 milliseconds. For domain adaptation, the team provided sentence-transformers snippets for fine-tuning.

Why it matters

We think this release solves a very real headache for edge RAG pipelines and on-device search. Multilingual retrieval has historically demanded either heavy, multi-gigabyte models or cloud API lock-in. Squeezing competitive 11-language retrieval into a 350M footprint means developers can now run local semantic search on modest hardware with zero external API fees.

The main caveat is document length: these models are explicitly tailored for short-context collections rather than long-form research papers. If your use case revolves around product listings, knowledge base snippets, or customer service tickets, they are well worth taking for a spin.