macOS 27 Turns the Mac Into a Local AI Hub With fm

X (Twitter) Facebook Pinterest LinkedIn E-mail

Apple has built fm into macOS 27, a Terminal tool that gives access to the language model Apple Intelligence runs on the Mac itself, and that can even spin up a local API compatible with Chat Completions — no account, no connection, and no per-token cost. As Appleismo explained, the move turns every Apple Silicon Mac running macOS 27 into its own inference point for simple tasks, which matters to both power users and companies looking to cut their dependence on the cloud.

The key facts about fm and Apple’s on-device AI in 20 seconds

macOS 27 ships with fm preinstalled, giving Terminal access to Apple’s local model.
fm serve exposes a Chat Completions-compatible API right on the Mac.
The model has roughly 3 billion parameters and a small context window.
Apps like LocalKeep already use the same model without sending data to any server.

The tool has gone viral on social media in recent days, sometimes with exaggerated claims. What’s confirmed is more modest, but it’s still worth a look for anyone thinking about AI architecture: a small, free model, already installed on any macOS 27-compatible Mac, built for summarizing, extracting, classifying, and rewriting.

From Terminal to a local API

fm offers seven commands: available, chat, count-tokens, license, respond, schema, and serve. fm respond fires off a query from scripts, --schema forces the model to return valid JSON, and fm serve starts an HTTP server on the machine itself with an interface compatible with OpenAI’s Chat Completions API. In the test published by BestHub, the server responded to /v1/models and /v1/chat/completions with the system model.

sudo fm license
fm schema object --name Ticket --string subject --string priority > ticket.json
cat email.txt | fm respond --instructions 'Classify this email' --schema ticket.json

For a technical team, that means tools already built for cloud models can point to an office Mac just by changing the service address, and repetitive tasks on internal data — like classifying emails or pulling fields out of documents — can be handled without the information ever leaving the machine. Structured output via JSON schemas is what makes it practical to plug into automated workflows.

There are important caveats. According to mac.install.guide, the local model’s context window is 4,096 tokens, though BestHub measured 8,192 on a Mac mini running macOS 27.0.1. The --model pcc option, which Apple showed at WWDC to route requests to a larger model on Private Cloud Compute, isn’t available in macOS 27.0. And on Apple’s developer forums, it’s still an open question whether a distributed app can call fm serve programmatically, since the tool’s legal notice restricts that unless explicitly permitted. Apple pointed to its license agreement without clarifying.

Apps already built on the local model

fm isn’t the only way to tap the model. Apple exposes it to developers through the Foundation Models framework, and apps are already built on it. One of them is LocalKeep, developed by David Carrero, editor of Revista Cloud, which summarizes, extracts, and rewrites text on iPhone, iPad, and Mac using the same model that powers Apple Intelligence. According to its website, it works without accounts, servers, or analytics, and also without a connection, and it integrates into the Share menu of apps like Safari, Notes, or Mail. It requires an iPhone 15 Pro or later, an iPad with M1 or later, or a Mac with Apple Silicon, with Apple Intelligence turned on, and it’s free at launch.

The approach illustrates the other side of fm. While the Terminal tool is built for people comfortable on the command line, apps built on Foundation Models bring the same capabilities to non-technical users, with the same privacy promise: processing happens on the device’s chip.

What it can and can’t do

Apple describes the local model as suited for summarizing, extracting information, classifying, rewriting, and short creative writing, and advises against using it for code, math, logical reasoning, or general-knowledge questions. In mac.install.guide’s tests, it invented a nonexistent Homebrew command, though it summarized and tagged text well. It doesn’t replace large cloud models, but it can absorb some of the routine work that’s currently billed by the API — a debate Revista Cloud covered in Pay Per Token or Buy a 24GB GPU?. Noticias.ai has also published a detailed guide to fm and its commands.

Frequently Asked Questions

What is fm in macOS 27?

It’s a Terminal tool preinstalled in macOS 27 that gives access to Apple Intelligence’s local language model, with no account, no connection, and no per-token cost.

Can fm be used as a local API?

Yes. The fm serve command starts a Chat Completions-compatible server on the Mac, though Apple hasn’t clarified whether distributed apps can use it programmatically.

What is LocalKeep?

An app for iPhone, iPad, and Mac that summarizes, extracts, and rewrites text using Apple’s local model, with no accounts or servers, according to its website.

What tasks is Apple’s local model good for?

Summarizing, extracting data, classifying, and rewriting. Apple advises against using it for code, math, or general-knowledge questions.

Sources: