I had Claude Code hand off big summarizing jobs to the on-device model in macOS 27

macOS 27 introduces 'fm', a command-line tool for Apple's on-device Foundation Model, which can summarize large texts locally. This allows users to reduce Claude Code token usage by having the local model summarize content like session transcripts and logs. The model handles inputs up to about 6,500 tokens, with a 2,500-token summary taking around 11.7 seconds. While good at faithful compression, its judgment quality is poor, often misinterpreting tasks or generating incorrect values, and it performs best when summarizing one coherent unit of text at a time.

This report uniquely details the practical limitations and performance characteristics of Apple's new on-device Foundation Model in macOS 27, unlike general announcements.