Liquid AI releases open model LFM2.5-24B-A2B for on-device SI agents: 24 billion parameters, 2 billion active per token
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
An agent model with open weights can be run on a laptop, not only in the cloud.
The mixture-of-experts model (64 experts per block, 4 selected), with a 131,000-token window and a 128,000-token vocabulary, calls tools and works through multi-step tasks in popular agent environments. The weights, including the base version, are open on Hugging Face and supported in llama.cpp, MLX and vLLM. According to the company, on the LiveClaw tool-use test it scored 48.09 against 38.19 for Qwen3.6-35B-A3B, and 96.75% on IFEval; speed is 184 tokens/s on Apple M5 Max and more than 21,000 output tokens/s on NVIDIA B300.
Sources 1