Liquid AI’s 24B model activates 2B parameters for local tool-use agents

What changed

Liquid AI’s new LFM2.5-24B-A2B uses a mixture-of-experts design: 24 billion parameters are available, while about 2 billion are active for a given token. The company positions it for local agents that need tool use and a long 131K context window, with open weights available through Hugging Face.

Liquid reports throughput of 184 tokens per second on an Apple M5 Max. Its release materials also highlight LiveClaw results against another open model, but performance depends on the benchmark and task. The model does not lead on every coding or knowledge measure, so the headline speed should not be read as a blanket quality win.

What the result means

The practical question is whether a developer can keep sensitive context on a device without giving up useful agent behavior. Sparse activation can lower inference work compared with a dense model of the same total size, but memory still has to hold model weights and local hardware changes the experience. The model card and evaluation tables are the place to compare before choosing it for a real workflow.