LFM2.5-230M Sets New Benchmark for Edge AI Data Extraction

LFM2.5-230M: performance on the edge

Liquid AI’s newest model, LFM2.5-230M, shatters expectations for on‑device language processing, delivering data‑extraction accuracy that beats models more than ↑4x its size. Built on the proprietary LFM2 framework, the 230‑million‑parameter foundation interleaves gated convolutions with grouped‑query attention, keeping memory under ↓400MB while supporting a 32K token context window.

Benchmarks released on the company blog show the model surpassing Alibaba’s Qwen3.5‑0.8B and Google’s Gemma 3 1B on extraction‑focused tests, scoring 43.26 on the BFCLv3 tool‑use suite versus 39.58 for IBM’s Granite 4.0‑350M. On a Snapdragon Gen4‑powered smartphone it reaches 213 tokens per second, and even a Raspberry Pi 5 manages 42 tps, proving the claim that it can run “anywhere.”

“We designed LFM2.5‑230M to empower developers to build autonomous edge pipelines without relying on costly cloud APIs,” the release read.

Enterprise licensing follows a dual‑use model: free for firms under $10 million annual revenue, with paid agreements for larger players. This mirrors a broader industry shift noted by Reuters toward efficient, locally‑deployed AI.

Why it matters for data‑heavy workloads

Traditional ETL scripts crumble when document layouts change. By embedding a lightweight extraction engine directly on device, companies can replace brittle rule‑sets with adaptive language models, cutting latency and cloud spend dramatically.