Artificial intelligence startup Liquid AI released the open-source LFM2.5-2.6B agent model yesterday local time. With just 2.6 billion parameters, LFM2.5-2.6B is highly suitable for local on-device deployment and can run even on smartphones, while still being capable of supporting agent workflows involving planning, tool calling, and multi-step processing.
ITHome learned that LFM2.5-2.6B was pretrained on a dataset containing 34T tokens and enhanced its support for non-Latin-script languages by expanding the existing tokenizer. Liquid AI is offering both base and post-trained versions of the model on the Hugging Face platform.
According to data provided by Liquid AI, when benchmarked alongside larger models such as gemma-4-E2B-it, gemma-4-E4B-it, Qwen3.5-4B, and Qwen3.5-9B, LFM2.5-2.6B led in all instruction-following tasks and nearly all tool-use tasks, falling behind Qwen3.5-9B only in the BFCLv4 test.
LFM2.5-2.6B comprehensively outperformed the two Gemma models on agent tasks and performed similarly to the Qwen models. In the STEM field, it led in the AA Omniscience test and fell behind Qwen3.5-9B only in the mathematics test. Programming was the only area in which the larger models maintained an advantage.