Pared Removes Apple Intelligence Models From macOS 27 With SIP Left On
Software / news
Pared Removes Apple Intelligence Models From macOS 27 With SIP Left On
Apple took the off switch out of macOS 27 while the models take 14 to 30GB. A 143-star MIT tool now deletes them through Apple's own asset service.
Pared, an MIT-licensed utility on GitHub, removes Apple Intelligence models from Macs running macOS 27 while leaving System Integrity Protection (SIP) switched on. It appeared on Hacker News's technology list on October 7, 2026, two weeks after reports that the models were taking up to 30GB of disk.
Why the tool exists
Apple, in macOS 27 and iOS 27, "removed the toggle to disable Apple Intelligence, so users unequivocally have to set aside space for Apple's models", MacRumors' Tim Hardwick wrote on September 23. Apple says the features need up to 14GB on Macs with an M3 chip or later and at least 12GB of memory, and up to 8GB on other compatible Macs.
Measured numbers run higher. The figures below are from the MacRumors article, which checked System Settings, General, Storage, and then the info button beside "macOS". Two are MacRumors' own, two are attributed to Ars Technica and two to Reddit users.
- Apple ceiling, other Macs8 GB
- Apple ceiling, M3 and newer14 GB
- M4 Pro Mac mini (MacRumors)20.67 GB
- M3 MacBook Air (Ars Technica)22.42 GB
- M3 Pro MacBook Pro (Reddit)24.35 GB
- Unspecified Mac (Reddit)30.16 GB
Source: MacRumors, September 23, 2026, citing its own measurement, Ars Technica and Reddit users
Ars Technica's M1 MacBook Air came in at about 14GB, so not every machine overshoots. The 20GB-plus readings are single machines, not a survey.
How Pared works
Per its README, Pared "uses Apple's asset service to remove models" and uses configuration profiles to block future downloads. You pick which features to keep, turn off or leave at the app's default. Writing Tools and Genmoji are among those listed. Choices are saved as a draft first, and model removal is a separate confirmation step.
The profile is not applied silently: the README says you must install the generated profile in System Settings to block future downloads. It requires Apple silicon and macOS 27 or newer.
| Install route | What the README says |
|---|---|
| App | Download and drag to Applications |
| Terminal wizard | A curl command that pipes install.sh to sh |
| Nix or Home Manager | Modules available; they remove models during activation by default |
| Source | Needs Xcode 27 or newer and tools/build-app.sh |
The terminal wizard is the route to slow down on. Piping a script from a GitHub release into a shell runs whatever that file contains on the day you run it. The app download and the source build are the paths where you can look first.
What it does not do
The README is candid about limits: "For features Pared can't download directly, use their Apple app." It points to a separate page on how profiles and model removal work for "implementation and compatibility limits", which was not read.
Three things the README does not state in the text read here: how much disk a typical removal frees, whether administrator rights are needed, and which release is current. No figure for space reclaimed is given, so the chart above is the ceiling of what a removal might return, not a measured result.
Stars are not adoption. The repository had 143 stars, 4 forks and 0 watchers when read, and a name behind the account, 4evy, that the page does not expand. Apple has said nothing in the sources read about whether removing models through its asset service is supported, or whether a later macOS 27 update would restore them.
For a Mac with a 256GB drive, a 14GB to 30GB allocation is the difference between a working update and a full disk. For earlier coverage of on-device assistants and where they send data, see our piece on Meta's Muse dispute; for another agent-era utility that spread fast on GitHub, see rea.
Pared's next test is the first macOS 27 point update. If Apple's installer restores the models, the profile is the only thing standing between a user and the download.
Sources
More in Software
- 01Orca Gains 1,590 Stars in a Day and Answers Three of crystl's Six tmux Limits, With 4,303 Open PRsThe MIT-licensed worktree app for coding agents has 98,318 stars and 4,306 open issues; its README addresses three of the six tmux limits crystl lists.
- 02OpenCode Has 217,837 Stars and a Security Policy That Says Its Permission Prompts Are Not a SandboxThe MIT-licensed coding agent's SECURITY.md tells users to run it in a Docker container or VM; its README labels the curl-to-bash install "YOLO".
- 03CLI-Anything Has 55,252 Stars and Three Test Totals, and Its Per-Harness Lines Sum to 2,622The HKUDS project that wraps desktop software for coding agents reports 2,461, 2,464 and 2,280 passing tests in one README, and its last push was Sept. 22.
- 04Linux 7.3 Ships With 18,765 Commits, a Filesystem That Fails Everything and Per-Cgroup BPF SchedulersLinus Torvalds released the kernel on Oct. 18 after a final week he called uneventful, with the biggest figures coming from storage and memory fixes.