Liquid AI's LFM2.5 Puts a 24B MoE on Your Device

Liquid AI has released LFM2.5-24B-A2B, a mixture-of-experts language model designed to run close to the user rather than in a distant data center. The model carries 24 billion total parameters but activates only about 2 billion per token, a design meant to deliver the knowledge capacity of a larger network while keeping the per-inference cost low enough for on-device deployment.

That tradeoff is the whole point. Mixture-of-experts architectures route each token through a small subset of specialized subnetworks, so a model can store far more information than it spends computing on any given step. For edge hardware — laptops, phones, and embedded systems where memory and power are tight — the gap between total and active parameters is what makes a 24B-class model practical at all.

Why it matters

Most capable open models still assume a GPU server is doing the work. Liquid AI has been building in the opposite direction, aiming its LFM line at local execution and agentic tasks that need to run without a round trip to the cloud.

• 24B total, ~2B active: broad capability with a small compute footprint per token
• Multilingual support: positioned for agents that operate across languages
• Edge-first: intended for on-device assistants and automation

As covered in Liquid AI's launch writeup, the model targets the growing class of edge agents — software that reasons and acts locally. The release ships under a custom license, so teams planning commercial use should read the terms before building on it.