✦
Summary
LFM2.5-24B-A2B is Liquid AI's reasoning model for fast, private agents running on local hardware. A hybrid architecture combines short convolutions, grouped-query attention, and sparse experts, activating about 2B of 24B parameters per token.
About LFM2.5-24B-A2B
Developers can run and fine-tune a local agent model while choosing between base and post-trained checkpoints. Core capabilities include 24B total parameters, Approximately 2B active parameters, Reasoning-focused output, and On-device tool use.
Public code or model weights are linked from the project, with setup instructions for running or adapting the system. Additional characteristics include Base and post-trained models, Local fine-tuning, Apple Silicon recipes, and GPU serving support.