LFM2.5 on-device: deciding an action is only the handoff
Liquid AI documents local agent integration; the application still needs to own permissions and tool execution.
A tool request is where the model hands work to the application. LFM2.5-24B-A2B is designed to reason and choose tools, but the surrounding agent software must receive those requests and execute them. I would make that boundary the starting point of a local-agent evaluation.
In an article published October 13, 2026, Raahul Vignesh and colleagues at Liquid AI describe an open-weight model that reasons before answering. They document support in llama.cpp, MLX, vLLM and SGLang, plus Hermes Agent configuration for a local server exposing an OpenAI-compatible API.
My architectural preference would keep argument validation and permissions in the executor. I would also retain the tool request and its result so that a poor decision can be distinguished from an execution failure.
Liquid AI's own comparison reports different strengths across workloads. The model does not lead every coding or knowledge benchmark. Those differences are useful for selecting evaluation tasks; the advertised decoding speed does not establish end-to-end agent reliability.
The operational choice still belongs to the application owner: which actions may proceed autonomously, and which require approval?