U.S. AI company Liquid AI unveiled the "LFM2.5-2.6B" on August 4, a model designed to run AI agents entirely on edge devices such as smartphones, and it is currently available for free download from platforms including Hugging Face. With just 2.69 billion parameters, it delivers performance on agent tasks such as tool calling and multi-step instruction following that rivals models four times its size, eliminating dependency on large cloud-based language models. According to Liquid AI's announcement, the model was trained on approximately 34 trillion tokens of data, featuring a vocabulary of 128,000 and a context length of 131,072 tokens, and it supports Japanese. Two variants are offered: "LFM2.5-2.6B-Base" for fine-tuning and "LFM2.5-2.6B" for agent workloads. The model scored 56.88 on the BFCLv4 benchmark, surpassing Google's Gemma 4 E28 it (5.18B) at 36.98 and Qwen3.5 (4.7B) at 50.56, approaching the 60.13 achieved by Qwen3.5-9B (9.7B). For CPU-based inference, it achieves decoding speeds of 220 tokens per second on Apple's M5 Max and 113 tokens per second on AMD's Ryzen AI Max+ 395, while processing 30 tokens per