Liquid AI's 230M model beats models four times its size at data extraction
Small models are quietly winning the jobs nobody brags about. Liquid AI says its tiniest LFM2.5 release yet — a 230-million-parameter model — outperforms rivals four times its size on data-extraction work, and it is small enough to run almost anywhere.
Liquid AI's new LFM2.5-230M tops models roughly four times its size on extraction benchmarks while running on hardware as modest as a phone, according to the company's announcement. The claim matters because extraction — pulling structured fields out of documents, tickets, and messy text — is one of the highest-volume jobs in production AI, and it rarely needs frontier-scale reasoning. A model this size fits in a few hundred megabytes of memory, which puts it inside smart-home hubs, point-of-sale terminals, and offline tools where calling a cloud API is slow, costly, or simply impossible.
The release rounds out a family that already leans edge-first. Liquid shipped DSpark draft-model acceleration for LFM2.5 earlier this month — we covered the technique when it landed alongside llama.cpp's own speedup work — and an encoder variant of the same 230M model has been up on Hugging Face since late July. The GGUF builds for llama.cpp have quietly collected around 140,000 downloads, evidence of real adoption rather than a demo.
There is a reason extraction is the beachhead. The task is narrow enough that a tiny model can be tuned to near-parity with much larger ones, and the volume is enormous: every invoice, receipt, support ticket, and form a company touches is an extraction job waiting to happen. Running those locally flips the economics — no per-token bill, no latency spike, and data that never leaves the device, which matters for finance and healthcare documents that compliance teams hate shipping to an API.
The take: the interesting competition in open models is shifting from who tops the chat leaderboards to who owns the boring, high-volume endpoints, and Liquid is building a portfolio for exactly that fight. A 230M model will not write your code, but it can read your paperwork all day on a battery.
What to watch: whether independent benchmarks back the "beats 4x-larger models" claim outside Liquid's own eval suite — small-model leaderboards have a habit of flattering their sponsors.
Running extraction at the edge — or still paying per-token for it? Tell us in the comments.
