Liquid AI, founded by former MIT computer scientists, today unveiled LFM2.5-230M, its smallest language model yet. The 230-million-parameter foundation model is built for on-device agentic workflows, targeting data extraction and local deployment on smartphones, laptops, and robotics.
According to Liquid, the model outperforms alternatives more than four times its size on selected benchmarks—specifically beating Alibaba's Qwen3.5-0.8B and Google's Gemma 3 1B at data extraction tasks. The company highlights that its small footprint enables operation "anywhere," as stated in the release blog post.
The model is released under a dual-use commercial license: free for individuals and companies with less than $10 million in annual revenue, with paid enterprise agreements required for larger corporations. This distinguishes LFM2.5-230M from other small AI models by leveraging the LFM2 architecture for high inference speeds.
This release signals a growing trend toward efficient, locally deployable AI models that reduce reliance on cloud infrastructure. For enterprises building autonomous edge systems or lightweight data extraction pipelines, such models offer cost-effective alternatives to larger, resource-intensive counterparts.
Liquid AI targets developers and engineers, focusing on practical deployment rather than raw parameter counts. The company did not disclose funding details or valuation in the announcement.