LiquidAI open-sources on-device agent reasoning model

Summary

LiquidAI released LFM2.5-24B-A2B, an open-weight mixture-of-experts reasoning model with 24B total and only 2B active parameters, built for fast, private agents running on-device. It thinks before answering, calls tools and handles multi-step tasks inside popular agent harnesses. On the company's own benchmarks it matches or exceeds larger MoE models such as Qwen3.6-35B-A3B on several tool-use and instruction-following tests, and it ships with day-one support for llama.cpp, MLX, vLLM and other inference tools.

Why it matters for our work

A reasoning agent that runs locally without external APIs widens private agent adoption in settings where data cannot leave the machine. Benchmark numbers are the company's own measurements, so real-world validation remains open.

Translated from the Korean original. Summaries may be translated and edited. Commentary reflects our perspective; forecasts remain the source’s views.