LFM2.5-24B-A2B: Built for fast on-device agents

You can now run sophisticated AI agents directly on your personal devices, with no reliance on cloud services for processing.

Liquid AI has released LFM2.5-24B-A2B, a 24 billion parameter language model that utilizes only 2 billion active parameters for each token. This architecture is designed for fast, private on-device agents. The model is competitive with larger models like Qwen3.6-35B-A3B on several agentic and tool-use benchmarks, outperforming it significantly on AutomationBench and LiveClaw. Its efficiency allows for rapid inference, achieving 184 tokens/s on an Apple M5 Max and up to 21K output tokens/s on an NVIDIA B300 GPU. LFM2.5-24B-A2B builds upon its predecessor with an expanded vocabulary, larger context window, and enhanced pre-training and reinforcement learning. The model, available in base and post-trained versions, is open-weight and accessible on Hugging Face. It is designed to excel at reasoning and tool-calling, with recommendations to pair it with retrieval or web search for knowledge-intensive tasks.

The development of LFM2.5-24B-A2B signifies a push towards highly efficient, on-device AI, directly addressing the need for privacy and speed in agentic applications. By activating only a fraction of its parameters, the model achieves performance comparable to much larger models while dramatically reducing computational requirements. This architectural innovation makes advanced AI capabilities accessible on consumer hardware, potentially democratizing the use of intelligent agents beyond cloud-based platforms.

This release challenges the current trend of scaling up models by focusing on parameter efficiency, paving the way for a new generation of specialized, lightweight AI. Developers and users can now expect more robust and responsive AI agents that operate locally, reducing latency and data transfer costs. The open-weight nature further encourages community development and fine-tuning, accelerating the adoption of on-device AI for a wide range of applications.