LiquidAI Launches 2.6B-Parameter Open-Source Model LFM2.5-2.6B, Runs Locally on Mobile Devices

On August 5, AI startup LiquidAI released LFM2.5-2.6B, an open-source agent model with 2.6 billion parameters capable of local deployment on edge devices, including smartphones. The model supports complex agent workflows such as planning, tool invocation, and multi-step processing. It was pretrained on a 34T-token dataset and features an extended tokenizer that enhances support for non-Latin scripts. LiquidAI has made both the base and post-trained versions of the model available on Hugging Face. Benchmark results show that LFM2.5-2.6B outperforms larger models—including gemma-4-E2B-it, gemma-4-E4B-it, Qwen3.5-4B, and Qwen3.5-9B—across all instruction-following tasks and nearly all tool-use evaluations. In agent-specific tasks, it consistently surpasses both Gemma models and performs comparably to the Qwen models. On the AAOmniscience benchmark in STEM domains, LFM2.5-2.6B leads overall, trailing only Qwen3.5-9B in mathematics. However, larger models still maintain an advantage in programming tasks.