LFM2.5 gives local agents a bounded lane

Liquid AI’s LFM2.5-2.6B is a 2.6B-parameter model aimed at on-device tool calling and multi-step workflows, with a 128K context window and support for llama.cpp, MLX, vLLM, SGLang, and ONNX. Liquid’s own evaluation says larger models still have a clear advantage on coding. Why it matters: Local inference is useful when the task is narrow enough to measure: sorting notes, extracting fields, or routing documents without sending them to a hosted model. It becomes a bad default when a team treats a small model’s tool use as proof that it can safely handle repository-scale work.

Try this: Run the same fixed extraction or tagging task locally with LFM2.5 and once with your current hosted model. Compare completion quality, tool-call failures, latency, memory use, and the review cleanup each version leaves behind.