Liquid AI Releases LFM2.5-24B-A2B On-Device Agent Model

The model is trained within popular agent harnesses, focusing on multi-step tool use rather than open chat. Base and post-trained weights are available on Hugging Face.

LFM2.5-24B-A2B features a 131,072-token context window and can run on laptops and mini PCs using Apple M5 Max or AMD Ryzen AI Max+ 395 processors via GGUF/MLX. It also supports single NVIDIA B300 for serving and offers FP8, GGUF, ONNX, and MLX 4/5/6/8-bit builds.

In performance tests, LFM2.5-24B-A2B leads or matches larger MoEs on five out of eight agentic and tool-use benchmarks. It achieved 27.69 on AutomationBench, more than triple the score of Qwen3.6-35B-A3B. However, it performed less strongly on SWE-Bench Verified (OpenCode) with a score of 51.00 compared to Qwen3.6-35B-A3B's 66.20.

The model is text-only and reasoning-only, built for local agent loops, with each assistant turn beginning with a <think> block. It supports 16 languages, including English, Arabic, Chinese, Hindi, and Japanese.

Its architecture retains LFM2’s 40-layer hybrid structure, incorporating 30 double-gated short convolution blocks and 10 grouped query attention (GQA) blocks. Each MoE block contains 64 experts with top-4 routing. Context increased from 32,768 to 131,072 tokens, and vocabulary doubled from 65K to 128K via an in-place tokenizer expansion for improved non-Latin tokenization.

Training involved scaling pre-training from 17T to 33T tokens, followed by mid-training with 2T tokens at a 32K context and then 300B tokens at a 131K context, focusing on tool calls and multi-turn interactions.