macOS 27 introduces fm command-line tool for Apple Intelligence
As macOS 27 (Golden Gate) brings a shift in local AI architecture by giving developers direct command-line access to Apple Intelligence. Rather than relying solely on proprietary cloud endpoints or restricted application wrappers, local Mac environments running Apple chips now expose an underlying large language model—known as Foundation Models—directly through the terminal via a new command-line tool named fm.
The Tech TL;DR:
- Direct CLI Access: macOS 27 introduces the
fmcommand-line utility, allowing direct interaction with local Apple Intelligence weights without web APIs. - Privacy Architecture: Processing runs entirely on local Apple Silicon hardware, ensuring zero data transmission to external cloud services and eliminating subscription costs.
- Developer Integration: The local model is accessible for custom scripts and automated workflows, though it trades off raw processing power compared to massive cloud models like ChatGPT or Gemini.
While Apple Intelligence typically routes generative tasks such as writing assistance, image creation, and third-party interactions through sandboxed application interfaces, the underlying infrastructure relies on on-device language models. According to reporting from Computer Bild, macOS 27 exposes this capability natively for command-line execution.
To initialize and interact with the hidden chatbot interface locally, users must operate on a compatible Mac equipped with an Apple Silicon chip and an active Apple Intelligence configuration. The activation sequence requires launching the terminal via Spotlight search, executing the license agreement protocol:
sudo fm licenseAfter confirming the prompt with “Yes” and entering the system administrator password, users can immediately engage the local model interface by running the primary chat command:
fm ChatArchitectural Trade-Offs: Local Latency vs. Cloud Capability
Operating a language model locally on an integrated hardware architecture drastically cuts round-trip network latency. However, benchmark evaluations show that these on-device Foundation Models carry a smaller parameter count than massive hyperscale cloud models such as OpenAI’s GPT-4 or Google’s Gemini. Consequently, complex reasoning tasks and extensive multi-turn logic may yield less nuanced outputs.
Despite these computational limits, the elimination of external data pipelines offers undeniable compliance advantages. For engineering teams and security analysts handling sensitive source code, proprietary enterprise data, or regulated datasets, running queries through an air-gapped local model prevents third-party telemetry leakage. Because Apple processes these instructions entirely on-device without enforcing recurring subscription fees, organizations can prototype local automation scripts and continuous integration pipelines without accumulating per-token cloud costs.
Implementation for Custom Automations
Beyond interactive chatting, the integration of the fm utility into shell scripts enables developers to bake local AI inference directly into local workflows.