Code & Chain · Signal Desk

Liquid AI Open-Sources LFM2.5-24B-A2B, Targeting On-Device Agents

2026-10-14AILatestLast updated: 13 Oct 2026, 11:36 pm UTC
Original sourceHugging FaceAdditional: Unite.AI

Summary

Liquid AI released open-weight model LFM2.5-24B-A2B, with 24 billion total parameters but only 2 billion activated per token, targeting continuous agent task execution on-device. Google says it achieves 184 tokens per second on Apple M5 Max and about 21 tokens per second output on supported GPUs, r…

Key points

  • Teams wanting to run agents locally without sending data to the cloud can download weights directly and deploy with existing inference engines.
  • Achieving usable agent reasoning with 2 billion activated parameters lowers the hardware barrier for private on-device agents.
  • Developers can deploy private agents on laptops or local GPUs, avoiding cloud costs and data leakage risks.
開源模型

Editorial note

This page is Code & Chain's editorial summary of public sources. It may be prepared with AI assistance and published through an automated workflow. Refer to the original sources; this content is not investment, legal, or tax advice.