Liquid AI's LFM2.5 targets fast on-device agents

A Hugging Face post introduces LFM2.5-24B-A2B, a model designed to run private, fast agents directly on devices.

Liquid AI has released LFM2.5-24B-A2B, a reasoning model designed to power fast, private AI agents that run directly on your own hardware.

A big model with a small footprint per token

LFM2.5-24B-A2B has 24 billion parameters in total, but only about 2 billion are active for each token it generates. That mixture-of-experts style design is the key to its pitch: the model can hold a lot of knowledge while keeping each step of reasoning cheap to compute.

Liquid AI says the model thinks before it answers, can call tools, and works through multi-step tasks inside popular agent frameworks. Because every reasoning token is inexpensive, the company argues you can leave agents running continuously on a laptop or workstation rather than paying for each step in the cloud. The Hugging Face post covers benchmark results, inference speed on both CPU and GPU, and instructions for plugging the model into existing agent harnesses.

Why it matters for builders

Agents are token-hungry. A single task can involve dozens of planning steps, tool calls and self-checks, and costs add up quickly with large hosted models. Small-active-parameter models change that math, making it practical to run background agents for things like inbox triage, code maintenance or monitoring without a growing API bill.

On-device execution also helps with privacy. Documents, code and personal data can stay on the user's machine, which is often a requirement for companies and for consumer apps handling sensitive information.

The trade-off is capability. A model with 2 billion active parameters will not match the best frontier models on the hardest problems. A sensible architecture is to let the local model handle routine steps and escalate difficult ones to a larger model when needed, with clear rules about what data can leave the device.

To try it, download the weights from Hugging Face and test it in a local runtime on a small set of your real agent tasks, measuring success rate and speed rather than relying only on published benchmarks.

You can find Hugging Face, Ollama and other local and hosted model platforms in our APIs & Model Platforms category.

Source: Hugging Face. This article is PickAI’s own summary and analysis based on the original reporting.