Liquid AI's 24B open model runs agents on a laptop

Liquid AI released LFM2.5-24B-A2B, an open-weight reasoning model built so that capable agents can run on hardware you already own. It is a mixture-of-experts with 24 billion total parameters but only about 2 billion active per token, which is what keeps it fast on a laptop. The design is a hybrid of gated short convolutions and grouped-query attention, with 64 experts and top-4 routing. Compared with the earlier LFM2 generation, the context window grows from 32K to 131K tokens and the vocabulary doubles to 128K, which mostly helps with non-Latin text. The model is reasoning-only, so it always produces a chain of thought before answering.

The pitch is agentic work rather than raw knowledge, and the benchmarks line up with that. Against Qwen3.6-35B-A3B, a larger open model, Liquid AI reports 27.69 versus 8.72 on AutomationBench, 80.71 versus 75.02 on ToolSandbox, and 96.75 versus 93.06 on IFEval for instruction following. It trails the bigger models where you would expect: 74.42 versus 83.69 on GPQA Diamond, and 51 versus 66.2 on SWE-Bench Verified for coding. Liquid AI is upfront about this and recommends pairing the model with retrieval or web search for knowledge-heavy tasks. On speed, it reports about 184 tokens per second on an Apple M5 Max and 114 on a Ryzen AI Max+ 395, with day-one support in llama.cpp, MLX, vLLM, SGLang, and ONNX.

Part of what produced these numbers is the training recipe. After supervised fine-tuning, Liquid AI ran on-policy distillation and then reinforcement learning inside real agent harnesses, including OpenClaw and OpenCode, rather than on static datasets. Both the instruct and base weights are on Hugging Face.

Why it matters

If you have wanted to run a tool-using agent locally for privacy or cost reasons, this is a concrete option that trades away some coding and knowledge strength for speed and strong instruction following. Test it on your own tool-calling workload before trusting the benchmark gap, and plan to add retrieval, because the model is deliberately light on stored knowledge.