Liquid AI has released the LFM2.5-2.6B model, a 2.6-billion parameter agent designed to run locally on edge devices with high efficiency and compatibility.
The model is trained using agentic reinforcement learning within popular harnesses to improve tool use and instruction following. It achieves inference speeds of 220 tok/s on an Apple M5 Max and 113 tok/s on an AMD Ryzen CPU while using under 2.5 GB of memory. Benchmarks show it competes with models four times its size, topping instruction-following metrics and matching larger Qwen models on agentic tasks.
The release aims to enable reliable on-device agents for high-volume workloads, available via Hugging Face and standard inference libraries.