Tiny AI Outperforms Giants on Key Tasks

Liquid AI, founded by former MIT computer scientists, has released its smallest AI language model to date, LFM2.5-230M. This 230-million-parameter model is specifically engineered for on-device agentic workflows, enabling efficient data extraction and local deployment on a wide range of devices including smartphones, laptops, and robotics.

Remarkably, LFM2.5-230M demonstrates superior performance in data extraction compared to models more than four times its size, such as Alibaba's Qwen3.5-0.8B and Google's Gemma 3 1B. This efficiency is achieved through its LFM2 architecture, which interleaves gated short-range convolutions with grouped-query attention, allowing for high inference speeds and a memory footprint under 400MB.

For enterprises, this model offers a cost-effective alternative to large, cloud-based AI solutions for tasks like 'AI ETL' and data parsing. It enables automation of repetitive processes at a fraction of the compute cost and latency, running directly on local hardware. The model is available under a dual-use commercial license, free for entities with less than $10 million in annual revenue, and requiring a paid enterprise agreement for larger corporations.

The LFM2.5-230M has also been demonstrated in advanced research, successfully translating complex natural language instructions into multi-step plans for a humanoid robot. Its immediate availability on Hugging Face and support across major inference ecosystems, including llama.cpp, MLX, and vLLM, ensures broad accessibility for developers and researchers.