Liquid AI has released LFM2.5-24B-A2B, a new open-weight reasoning model designed specifically for on-device AI agents running on laptops and PCs. The 24-parameter Mixture-of-Experts model activates approximately two billion parameters per token and serves as the final and largest member of the LFM2.5 family. Trained inside popular agent harnesses to target multi-step tool use, the model features base and post-trained weights now available on Hugging Face.
The architecture maintains a forty-layer hybrid structure featuring thirty double-gated short convolution blocks and ten grouped query attention blocks, while expanding its context window to 131,072 tokens and doubling its vocabulary size to 128,000. Liquid AI trained the model on thirty-three trillion tokens, incorporating agentic reinforcement learning within harnesses like OpenClaw and Hermes Agent. On consumer hardware, the model achieves 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395 using quantized local execution.
In benchmark evaluations, LFM2.5-24B-A2B matches or leads larger models on tool-use and instruction-following tasks, scoring 27.69 on AutomationBench and 96.75 on IFEval, though it trails on knowledge and coding benchmarks like SWE-Bench Verified. Day-one support spans platforms including vLLM, SGLang, llama.cpp, and MLX, providing developers with a fast and efficient local tool-calling solution. This release matters because it delivers high-performance, cost-effective local reasoning capabilities optimized for consumer hardware and multi-step agentic workflows.