LFM2.5-24B-A2B is Liquid AI's post-trained mixture-of-experts reasoning model, described on its card as the largest and final model in the LFM2.5 family and built for fast, private agents on-device. It has 24.0B total and 2.45B active parameters, 40 layers (30 double-gated short convolution blocks plus 10 grouped-query attention layers), 64 experts per MoE block with top-4 routing, a 128,000-token vocabulary and a 131,072-token context window, with 36 trillion training tokens reported and support for 16 languages. It keeps the LFM2-24B-A2B architecture but extends context from 32K to 131K and doubles the vocabulary from 65K to 128K. Unlike LFM2-24B-A2B it always thinks before answering, with its chat template opening a think tag, and it writes Pythonic function calls by default. Post-training combined supervised fine-tuning, on-policy distillation and agentic reinforcement learning inside the OpenClaw, Hermes Agent, Pi and OpenCode harnesses. On the card's 16-benchmark comparison against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning (30B-A3B), it scores 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B, 27.69 on AutomationBench versus 8.72, and 96.75 on IFEval versus 93.06, but trails Qwen3.6-35B-A3B on SWE-Bench Verified with OpenCode (51.00 vs 66.20) and GPQA Diamond (74.42 vs 83.69). Liquid AI reports 184 tokens/s decode on an Apple M5 Max and 114 on an AMD Ryzen AI Max+ 395 with llama.cpp, Q4_K_M and a 4K-token input, and over 21K output tokens/s on one NVIDIA B300 with SGLang, BF16 and high concurrency. The card advises against relying on it alone for knowledge-heavy tasks and suggests pairing it with retrieval or web search. Weights are released under the LFM Open License v1.0, which is not an OSI license: commercial use is licensed only while the user's annual revenue stays below 10 million US dollars, with an exemption for qualified non-profits doing non-commercial or research work. Separate repositories offer a Base model, FP8, GGUF, ONNX and MLX 4-bit builds.