Liquid AI ships a 230M on-device model

Liquid AI released LFM2.5-230M on June 25, its smallest model yet at 230 million parameters. It is built on the company’s LFM2 architecture, pre-trained on 19 trillion tokens with a 32K context window, and the weights are on Hugging Face for anyone to download, fine-tune, and run.

The point of the model is where it runs. Liquid reports 213 tokens per second on a Samsung Galaxy S25 Ultra and 42 tokens per second on a Raspberry Pi 5, fast enough for on-device agents and extraction pipelines that never touch a cloud API. On its own benchmarks the company says the model competes with, and sometimes beats, models more than twice its size on knowledge, instruction following, data extraction, and tool use. Liquid is also blunt about the limits: do not point it at heavy reasoning, math, or code generation.

What it means for you

A 230M model that runs on a phone or a Pi is not going to replace a frontier model, and Liquid is not pretending it will. The use case is narrow and real: pull structured fields out of forms, drive a small tool-using agent, or run text tasks on hardware you already own, with no per-token bill and nothing leaving the device.

Treat the benchmark claims as the vendor’s own until independent tests land. For the wider question of when a small open model is genuinely good enough, see the open coding models from Ornith, the GLM-5.2 open-weights results, and the basics in what an LLM actually is.