tl;dr: Liquid AI released LFM2.5-24B-A2B, a 24B-parameter MoE reasoning model with only 2B active parameters, designed for on-device agents with 131K context and open weights on Hugging Face. It hits 184 tokens/s on an Apple M5 Max and matches or beats Qwen3.6-35B-A3B on five of eight agentic/tool-use benchmarks despite having ~50% fewer parameters, though larger models still lead on knowledge and coding tasks. It ships day-one with llama.cpp, MLX, vLLM, SGLang, and ONNX support.