Liquid AI Releases LFM2.5-24B-A2B: A 2B-Active MoE for On-Device Agents

Liquid AI has released LFM2.5-24B-A2B, an open-weight Mixture-of-Experts reasoning model with 24 billion total parameters but only 2 billion active per token. The architecture is designed specifically for on-device agent workloads, where memory and compute constraints make dense models impractical.

The model scored 48.09 on the LiveClaw benchmark and achieved a decode speed of 184 tokens per second on an Apple M5 Max, demonstrating that MoE designs can deliver strong reasoning performance while keeping active compute low enough for local hardware.

The release is notable for targeting the growing category of local agents — AI systems that run entirely on consumer or edge devices without cloud round-trips. By keeping active parameters at 2B, Liquid AI aims to make agentic reasoning feasible on laptops and workstations rather than requiring data-center GPUs.

Open-weight distribution lowers the barrier for developers and researchers to experiment with MoE-based agent architectures. Whether LFM2.5-24B-A2B can match larger dense models on complex multi-step tasks will depend on real-world agent benchmarks beyond LiveClaw.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.