LFM2.5-24B-A2B runs private AI agents on-device

Liquid AI has published a post on Hugging Face about LFM2.5-24B-A2B, a model it positions for running agents directly on a device, quickly and privately. Judging by the name, it has 24 billion parameters in total, with roughly 2 billion active when generating a response, a design that typically saves compute.

This matters because an agent calls the model many times, not once: it reads documents, invokes tools and plans steps. In the cloud, each such chain adds latency and cost, and the data leaves the company's perimeter. Running locally removes part of that problem, especially for financial, HR and contract information. Specific figures on quality and hardware requirements are best checked in the original post and tested on your own tasks.

Write down which data in your company should never leave the building: bank movements, payroll, contracts, the customer base. The 1Cifer platform is built on the same principle: each role gets its own agent, and access to 1C, CRM and bank data follows the company structure inside a protected perimeter. Only with that list in hand does it make sense to decide what to run locally and what can safely go to the cloud.