Deploy local agents everywhere with LFM2.5-2.6B

LFM2.5-2.6B is a small, efficient language model designed to power capable agents entirely on-device, such as laptops and phones. It is built using a four-stage post-training process that includes supervised fine-tuning, teacher specialization, distillation, and a novel Agentic Reinforcement Learning (Agentic RL) phase. Despite its 2.6B parameter size, it performs competitively against models four times larger on benchmarks for tool use, instruction following, and agentic tasks. The model is optimized for efficient inference, achieving high token-per-second rates on both CPUs and GPUs, making it suitable for deploying agents at scale while maintaining data privacy.