liquid ai ships two tiny search models beating rivals twice their size

liquid ai shipped two models perfect for fast, accurate search across languages, running on almost any hardware

what they released: two tiny retrieval models (lfm2.5-embedding-350m and lfm2.5-colbert-350m) built to search product catalogs, faqs, support docs, and knowledge bases quickly and across languages

why it matters for business:

• finds what people mean, not just what they type – a search for "laptop won't turn on" still surfaces your doc titled "notebook fails to boot." fewer dead-end searches and support tickets than old keyword search

• multilingual out of the box – someone can search in french and find your english product listing. 11 languages covered, no separate model per market

• fast and cheap – query results in under 2ms on a server, under 10ms on a laptop or edge device. small enough to run almost anywhere at near-zero cost

• small but beats bigger – on multilingual and cross-language retrieval benchmarks they top qwen3-embedding-0.6b (nearly 2x the size) plus established models from alibaba, baai, and lighton.

the two flavors, in plain terms:

- embedding model = cheapest and fastest, smallest storage footprint. pick this when you want speed and low cost

- colbert model = more accurate, matches search terms word-by-word, costs more storage. pick this when getting the right result matters more than saving space

labs are starting to ship models built for narrow, specific business jobs-to-be-done rather than one giant do-everything model – and this is a clean example

you can see how the colbert model works below in the smart tool selection demo (by thehype): there's a list of 151 tools, and instead of scrolling to find the right one, you just type what you need in plain language and the model instantly surfaces the 5 most relevant