Liquid AI’s LFM2.5-230M is a 230‑million‑parameter text‑only model built for on‑device agentic work such as data extraction and simple tool calling. Its hybrid architecture mixes LIV convolution blocks with grouped‑query attention, giving fast CPU inference: about 213 tokens per second on a Galaxy S25 Ultra and 42 tokens per second on a Raspberry Pi 5, with a memory footprint of roughly 300 MB in 4‑bit form. The model ships as an open‑weight checkpoint on Hugging Face and is ready to run in llama.cpp, MLX, vLLM, SGLang, or ONNX.
For teams that need to pull structured information from large volumes of documents—like parsing clinical reports into fields—LFM2.5‑230M can run locally on commodity hardware, eliminating per‑token API costs and keeping data private. Its instruction‑following scores (71.7 on IFEval, 38.4 on IFBench) beat comparable models up to four times its size, while its data‑extraction result on CaseReportBench (22.5) shows solid performance for focused tasks.
The model is not meant for heavy reasoning, advanced math, code generation, or creative writing. If those capabilities are required, a larger model should be chosen instead. LFM2.5‑230M excels when the goal is to turn a natural‑language request into a quick tool call—such as checking a candidate status, setting a thermostat, or selecting a skill on a robot—because it reliably outputs a Pythonic function call between special tokens that can be executed and then turned into a plain‑text answer.
Developers can start with the provided Transformers example: load the model with bfloat16, apply a chat template, generate with temperature 0.1, top_k 50, repetition_penalty 1.05, and do_sample=True. Fine‑tuning recipes (SFT, DPO, GRPO with LoRA) are available as Colab notebooks via Unsloth and TRL for further specialization.
In short, if your product needs low‑latency, on‑device extraction or simple tool use without relying on cloud APIs, LFM2.5‑230M offers a practical, lightweight solution that fits on phones, single‑board computers, and edge robots.
#AI #MachineLearning #EdgeComputing #ModelDeployment #DataExtraction #ToolUse