Mac Terminal Gets a Free LLM Squeezed In, and the 4096-Token Quota Is a Bit Jarring

macOS 27's Terminal now includes an fm command. According to a tweet, every Mac upgraded to macOS 27 comes with a free LLM: no account, no API key, no internet connection, and no token bill.

Setup isn't complicated. First run sudo fm license and enter your Mac password to accept the license; then use fm available to check whether the model has already been downloaded; after that, fm respond 'your question' lets you ask a question. The tweet says any Apple Silicon Mac can use it, including M1.

It's useful for more than Q&A. You can pipe files into it, for example cat notes.txt | fm respond -i 'summarize this'; you can have it return JSON according to a schema; fm serve spins up an OpenAI-compatible API on localhost; and sessions are saved in ~/.fm/sessions/. For developers, the last two points are more useful than chatting in the terminal.

The limitations are clear too. Each session is capped at 4096 tokens, shared by the prompt, instructions, and response. The model itself is small, so don't expect it to write complex code or solve math problems. It's suited to summarization, extraction, classification, and rewriting. The tweet explicitly suggests skipping code and math.

The community reaction is more interesting than the feature introduction. In the replies, serx says 4096 tokens also has to cover the response, which is too tight. Deep thinks it can only be used for short tasks because the model is too small. Helge Rebhan complains that it's yet another 5GB piece of bloatware, just for basic things like summarizing email, and that it can't be uninstalled. SurfingTheUniverse tried asking what /usr/bin/fm is, and the model said it was a tool for managing the file system; fm --help turned out to be more useful.

Others pointed out usage issues. Toaster Tom says you don't have to keep using fm respond; you can just run fm chat to enter a normal conversation. Kristoffer Laurson asks why use a clunky CLI when local model tools like LM Studio have had usable interfaces for years. Jack Chiang asks whether it's available in China, and dell reminds him that it's limited to macOS 27.

The value of this news isn't that 'Apple finally has an LLM.' Local model tools have existed for a long time; Ollama, LM Studio, and nightsmith can all run offline. What Apple has done is provide a system-level fallback: no configuration, offline by default, clear privacy boundaries, and readily available in the terminal. It's more like a short-text processor bundled with the system than a replacement for local LLM workflows.

The hard 4096-token limit means it can't handle heavy work. But for summarization, classification, rewriting, and extracting fields from private files, this 'already on the machine' convenience has real value—provided you accept that it isn't very smart and that it takes up disk space.