Liquid AI Drops a 24B Local Agent Model With Only 2B Active Parameters
LFM2.5-24B-A2B is downloadable now with a longer context window and tool-use focus; its speed and benchmark gains are Liquid AI's measurements.
A larger model built for local agents
Liquid AI released LFM2.5-24B-A2B on October 13, pitching a downloadable reasoning model for agents that call tools without sending their work to a cloud API. Its mixture-of-experts design has 24 billion total parameters but activates about 2 billion for each token. Both the base and post-trained weights are available on Hugging Face. The model card lists an LFM1.0 license, so open weights should not be mistaken for a standard permissive open-source license. Liquid AI documents local setups through tools including llama.cpp, MLX, vLLM and SGLang.
What the release claims
The company says the context window grows from 32,768 to 131,072 tokens compared with the previous LFM2-24B-A2B and that the new model always generates a reasoning trace before its answer. In Liquid AI’s tests, it led or matched larger comparison models on five of eight agentic and tool-use benchmarks. The company also reports fast decoding on an Apple M5 Max. Those are its own evaluation conditions, not an independent guarantee for every device or workload. Liquid AI says larger competitors retain an advantage on some knowledge-heavy and coding tests, and recommends pairing this model with retrieval when factual knowledge matters.