macOS 27 ships with Apple Intelligence's on-device models, and according to a new open-source tool called RemoveMacAI, the system no longer has a single switch to turn the feature set off, and the models stay on disk even after you disable the features. The project's pitch is one command: turn off Apple Intelligence, remove its downloaded models, and stop macOS from downloading them again.
The repository passed 519 stars within days, and the Hacker News thread hit 367 points and 226 comments on October 5, 2026. Most of the discussion was less about the tool than about what it says about the operating system. This post covers what RemoveMacAI changes, how it works under the hood, the security questions raised about its installer, and what to do if you would rather not use a third-party tool.
TL;DR: the questions people are asking
| Question | Short answer |
|---|---|
| What does it do? | Disables Apple Intelligence features, deletes the models, blocks re-download. |
| What does it need? | Apple silicon and macOS 27 (tested on 27.0; use v0.2.3+ on 27.0.1). Not supported on 26 or earlier. |
| How much space back? | Commenters cite roughly 12-14 GB; run removemacai status for your actual number. |
| Is it reversible? | Yes: removemacai revert. macOS downloads models again when a feature needs them. |
Does it disable SIP or edit /System? | No. System Integrity Protection stays on and nothing under /System is modified directly. |
| Is the installer safe? | Releases are attested, but the one-line script is not. Read it or verify the release. |
| Does dictation still work? | Yes, per the FAQ. |
What macOS 27 changed
The project describes two problems. First, macOS 27 has no single Apple Intelligence toggle. Second, the models stay on disk after you switch features off. In the Hacker News thread, users cited 12 to 14 GB, with one saying Storage settings listed about 30 GB, and several said that on a 256 GB Mac that is a meaningful share of the drive.
Nuance matters here. A commenter who explicitly enabled the new Siri-powered search said it was opt-in for them. Another said the models were downloaded at install time and sat on disk despite following every available disable step. Both can be true on different machines or configurations, so check your own Storage pane before assuming either.
For readers following the platform side, our coverage of Apple's September 9 AI announcements and the iOS 27 Siri model delegation code explains why these on-device models exist in the first place.
What RemoveMacAI turns off and removes
Features turned off, per the README: Siri (including "Hey Siri" and the menu bar icon), Writing Tools, Genmoji, Image Playground, the ChatGPT extension, summaries in Mail, Messages, Safari, Notes and notifications, Mail smart replies, inline text predictions, Spatial Photos, Photos Clean Up and Xcode predictive code completion.
Models removed: the Apple Intelligence foundation models, plus the models for image generation and Genmoji, Spatial Photos, Photos Clean Up and Xcode code completion.
What stops working afterwards: the features above, apps that rely on Apple's on-device models (the Foundation Models framework and the Use Model action in Shortcuts), Visual Intelligence and natural-language editing in Calendar. If you build apps on Apple's on-device model API, test before removing, because your own app can break.
How it works
RemoveMacAI does three things, none of which require turning off System Integrity Protection:
- A configuration profile applies Apple's own restriction keys for Apple Intelligence and forces settings that have no restriction key. macOS requires you to approve the profile in System Settings, so the tool cannot install it silently.
- Model removal through Apple's asset service, rather than deleting files under
/System. - A download block. The profile redirects each removed model's download to a closed local port, so macOS does not fetch it again.
Removing the profile restores previous settings. The tool says it makes no network requests and collects no data. It builds on an earlier tool called pared, which first mapped the asset service and several settings keys, and credits it in the repository's third-party notices.
Install, usage and revert
The README offers a one-line installer and a Homebrew tap. Core commands:
| Command | What it does |
|---|---|
removemacai | Show the current state, then turn Apple Intelligence off |
removemacai status | Show each feature and the size of the models on disk |
removemacai off --keep <features> | Leave listed features on |
removemacai off --dry-run | Show changes without applying them |
removemacai features | List names accepted by --keep |
removemacai revert | Undo all changes |
Start with status and off --dry-run. If you only dislike some features, use --keep for the ones you use. After removal, System Settings, then General, then Storage may keep counting the models until macOS deletes the files on its own schedule, even though the asset service has released them, according to the FAQ.
The security argument: should you curl-pipe this?
The most heated half of the thread was not about Apple at all. The installer is curl ... | bash, and commenters split three ways.
- Defenders said the script and the binary come from the same repository and author, so reading the script adds little if you already trust the author. The script downloads the latest release, verifies a SHA-256 checksum and runs from a temporary directory, installing nothing permanently.
- Critics pointed out that a server can serve different content to a piped request, that checksums fetched from the same place add little against a compromised host, and that a script that may call
sudodeserves a read first. - A sharper technical point: the project advertises build provenance attestations, which is a real strength, since releases are built from tags by GitHub Actions and can be checked with
gh attestation verify removemacai-darwin-arm64.tar.gz -R omlahore/RemoveMacAI. But a commenter noted the install script is referenced from the mutablemainbranch and is not covered by the attestation, so piping it does not verify anything by itself.
One thing is certain: the tool needs your approval for a profile and touches system state, so treat it like any administrative software. A safer sequence is to download the release, run the attestation check, read the script, and then run the binary.
Related reading on the broader risk: our write-up of supply chain security scanners for agents, and the Strata coverage where the same curl-pipe debate resurfaced the same day.
Without a third-party tool
If you do not want to run an installer, you have fewer options, and they are partial:
- Turn features off individually in System Settings and in each app. This leaves the models on disk.
- Read a manual walkthrough, which commenters linked from Apple community documentation. It describes steps comparable to what the tool automates.
- Build or download the configuration profile yourself and inspect it. One commenter noted the tool largely generates a profile on the fly and suggested shipping a pre-built
mobileconfiginstead, which is easier to audit. - Delay the upgrade. Some commenters are staying on macOS 26 until a master switch appears. That carries a security tradeoff, since you stop receiving the newest fixes.
What this means for builders
For developers, the practical takeaways are different from the consumer debate:
- On-device AI is now a platform default. If you ship Mac apps that call the Foundation Models framework, some users will have removed the models, so check availability and degrade gracefully.
- Disk budgeting for local AI is real. The same week, Strata asks for about 80 GB to run a 125B local model. Users who reclaim space from system models often spend it on models they chose, which is a useful consent contrast.
- Consent and defaults drive backlash. Commenters were more upset by the lack of a master switch and a visible way to reclaim space than by the models existing.
- Check macOS 27 minor versions. The project notes behavior changed between 27.0 and 27.0.1, so pin the tool version.
Why this keeps happening: storage, soldered SSDs and consent
A large share of the thread was about hardware economics rather than software. Apple sells base Macs with 256 GB or 512 GB of storage, the SSD controller is integrated into the M-series chip, and upgrades cost far more than commodity drives. In that context, a 12 GB feature that you cannot easily remove reads as a cost transferred to the user. Defenders replied that the new Siri-powered search is convenient and that most buyers never touch these settings, so a default-on design serves the majority.
Both positions point at the same fix, which commenters kept repeating: a visible master switch, and a visible way to reclaim the space. Until Apple ships one, a profile-based tool like this is the only complete route, and its popularity is the signal.
Related reading
- Every AI feature Apple announced on September 9
- Apple's Siri can swap in Claude or ChatGPT: what the iOS 27 code shows
- Apple M6 Mac mini and M5 Ultra Mac Studio launch
- Apple's 1-trillion-parameter model on Mac Studio
- MacBook vs dedicated GPU for local LLMs
- Run a 125B model on a 12 GB GPU with Strata
- Supply chain security scanner for agent workflows
Primary: the RemoveMacAI GitHub README and release notes (v0.2.3) · the October 5, 2026 Hacker News discussion
Details are accurate as of October 5, 2026. RemoveMacAI is a third-party tool, not affiliated with Apple. Model sizes vary by Mac and configuration; run the status command and read the script before installing.
