Apple appears to have made its Foundation Models accessible directly from Terminal in macOS 27.
The interesting part is how little setup is required:
``` fm respond "Your Question" ```
You can also pipe text into it, request structured JSON output with `--schema`, and expose a local endpoint using `fm serve`.
On-device inference reportedly works on M1 Macs with 8GB RAM, with a 4K context window. There's also a Private Cloud Compute option for larger requests.
For me, the interesting use case isn't replacing Claude or ChatGPT. It's using a built-in local model for everyday shell automation: extracting information from logs, classifying files, summarizing documents, and generating structured data without API costs.
Curious whether anyone here has benchmarked it against Ollama or llama.cpp for these smaller tasks.