Liquid AI's smallest model yet LFM2.5-230M beats models 4X its size at data extraction, can run 'anywhere'

Insight summary

• •Liquid AI released LFM2.5-230M, a 230-million-parameter AI model designed for on-device workflows on smartphones, laptops, and robotics.
• •The model outperforms models over 4 times its size in data extraction tasks, including Alibaba's Qwen3.5-0.8B and Google's Gemma 3 1B.
• •LFM2.5-230M uses a hybrid architecture combining gated short-range convolutions and grouped-query attention for efficiency with a 32K context window.
• •It maintains a memory footprint under 400MB and achieves high decode speeds up to 213 tokens per second on a Samsung Galaxy S25 Ultra.
• •The model targets lightweight data extraction pipelines and autonomous edge systems, offering a cost-effective alternative to larger AI models.
• •Liquid AI provides the model free for companies with less than $10 million annual revenue under a dual-use commercial license; enterprises require paid agreements.
• •The model supports AI ETL processes for dynamic data extraction, replacing brittle rule-based systems and reducing reliance on costly cloud APIs.