A model with 230 million parameters just embarrassed competitors packing four times the weight. Liquid AI, the MIT spinout valued at roughly $2 billion, released its LFM2.5-230M foundation model on June 25, targeting on-device AI workflows where cloud access is either impractical or unwanted.
The headline number: LFM2.5-230M scored 22.51 on data extraction tasks using the CaseReportBench dataset. Alibaba’s Qwen3.5-0.8B, a model with 800 million parameters, managed 13.83. Google’s Gemma 3 1B, sitting at a full billion parameters, scraped together 2.28.
Small model, big implications for edge computing
The LFM2.5-230M was pre-trained on 19 trillion tokens and designed to run, as Liquid AI puts it, nearly “anywhere.” On a Samsung Galaxy S25 Ultra, the model clocks 213 tokens per second. On a Raspberry Pi 5, a single-board computer that costs less than a nice dinner, it still manages 42 tokens per second.
The pre-trained weights are already available on Hugging Face, meaning developers can start building with it immediately.
