Liquid AI releases LFM2.5-24B-A2B, a 2B-active sparse model for on-device agents

Liquid AI's latest model pushes the sparse-mixture approach toward phones and laptops. LFM2.5-24B-A2B carries 24 billion total parameters but activates only about 2 billion per token. That gives it the knowledge capacity of a mid-size model with the compute cost of a small one. The release is published on the Hugging Face blog and aims at agent workloads like tool calling, routing and multi-step tasks that must stay local for privacy or latency.

The 'A2B' design means memory, not compute, is the main constraint. All 24B parameters must fit in RAM, but each forward pass is cheap, so throughput on consumer hardware can be high once the model loads. Liquid has built its brand on architectures that differ from standard transformers to optimize edge efficiency.

The model enters a busy edge-AI week. DeepMind and Singapore shipped an on-device Gemma scam detector, NVIDIA's Nemotron 3.5 Lightning offers a 30B/3B-active open MoE, llama.cpp gained a Windows app, and EmbeddingGemma 2 was repurposed for speech-to-text. Small-active-parameter models are becoming the default design for private agents.

Watch for independent benchmarks against Nemotron 3.5 Lightning and Qwen small MoEs, and for license terms that determine commercial on-device use.