Liquid AI releases LFM2.5-24B-A2B, a 24B MoE with 2.45B active for local agents
In one sentence On October 13 Liquid AI published the weights of LFM2.5-24B-A2B, the largest LFM2.5 model: a 24-billion-parameter MoE with 2.45 billion active, reasoning-only, built for tool-calling agents running on laptops and workstations, with GGUF, MLX, ONNX and FP8 builds.
Liquid AI builds models with a hybrid architecture of convolution and attention blocks, optimized to run on local devices. Its new LFM2.5-24B-A2B is the largest in the family: 24 billion parameters in total, but only about 2.5 billion at work for each generated word, which makes it fast even on a laptop.
The model is designed for agents: it always reasons before answering and was trained to use external tools such as APIs or commands. Liquid itself warns that on its own it is not suited to knowledge-heavy questions and recommends pairing it with web search or documents.
The weights are on Hugging Face under the lfm1.0 license, which according to press coverage allows free and commercial use up to 10 million dollars of annual revenue. Ready-made builds already exist for llama.cpp, LM Studio, Ollama and Apple Silicon Macs.
Companies
Liquid AI
Tools
LFM2.5-24B-A2B
Tags
Sources