What is available
Liquid AI announced LFM2.5-24B-A2B on October 13. It is a text-only model built for reasoning and tool use. Its detailed model card lists 24 billion total parameters and 2.45 billion active parameters, with a context window of up to 131,072 tokens.
Official weights and compressed versions are listed on Hugging Face. The Q4_K_M file alone is 14.5 GB; that is a download size, not the complete memory requirement for running the model.
Read the speed result carefully
Liquid reports 184 generated tokens per second on an Apple M5 Max using llama.cpp, Q4_K_M and a 4K-token input. This is the developer’s measurement under those conditions, not our test or a speed promise for every Mac. Liquid also advises against relying on the model alone for knowledge-heavy tasks.
Check the license
The repository uses the custom LFM Open License v1.0. Commercial permissions include a $10 million annual-revenue threshold; redistribution and attribution conditions also apply. Read the linked terms for your intended use. Downloadable weights do not mean unrestricted use.
Our take for AI geeks
Editorial analysis: consider a narrowly scoped local experiment after checking your hardware’s capacity. Use invented data and tools limited to reading, then assess correctness, memory use and time to a finished answer. Running the model on your own computer can avoid a hosted model’s per-request charge, but hardware, power and maintenance still cost money. Serving many visitors still requires capacity. We have not downloaded or tested this model.