Liquid AI Releases LFM2.5-2.6B for On-Device Deployment on Smartphones
Liquid AI has released LFM2.5-2.6B, a 2.6-billion-parameter open-source model designed for local deployment on smartphones. Pre-trained on a 34-trillion token dataset, the model is optimized for agentic workflows like planning, tool use, and multi-step processing. This release advances the trend of on-device AI by proving that a highly compact 2.6B model can outperform much larger models in instruction following and tool use. It enables complex, agentic AI workflows to run locally on consumer hardware without relying on cloud APIs. LFM2.5-2.6B outperformed Gemma and Qwen models of larger sizes in instruction following and agentic tasks, though it lagged behind Qwen3.5-9B in coding and the Berkeley Function Calling Leaderboard (BFCLv4) benchmark. The model is available on Hugging Face in both base and post-trained versions, featuring enhanced support for non-Latin languages.
## BACKGROUND
Liquid Foundation Models (LFMs) are a class of generative AI architectures developed by Liquid AI, built with computational units rooted in the theory of dynamical systems to optimize for fast inference and real-world constraints. Benchmarks like the Berkeley Function Calling Leaderboard (BFCL) are used to evaluate how effectively these models can interact with external tools and APIs to perform complex tasks.