Run fast, private agents on-device with LFM2.5-24B-A2B

A new family of AI models, the Liquid Foundation Models (LFM), enables powerful and private agents to run directly on personal devices without needing cloud access. Built on a novel Liquid Mixture of Experts (LMoE) architecture, these models dynamically allocate computational resources to achieve remarkable speed and efficiency. The flagship 24-billion parameter model, LFM2.5-24B-A2B, was trained on an immense 15 trillion tokens of data and outperforms many larger models like Llama 3 70B. Specifically designed as an "agent-to-be" (A2B), it provides a powerful foundation for developers to create specialized, fast, and private AI assistants.