An open-source RemoveMacAI tool for macOS switches off Apple Intelligence and deletes Apple’s AI models, potentially freeing up over 12GB of storage.
RemoveMacAI is an open-source command line utility with one job: locate and delete the local Apple Intelligence assets on macOS 27, covering the on-device foundation model, image generation models and supporting language resources, roughly 12GB in total. Users run a single command with administrator privileges; the project is hosted on GitHub under an open-source license, so the code is auditable, and community discussion has focused on how far the deletion should reach and how it coexists with System Integrity Protection.
The change that created the demand began with macOS 27. Through macOS 26, Settings offered a toggle: switching Apple Intelligence off stopped download
Apple Intelligence was announced at WWDC in June 2024 and shipped that October with iOS 18.1 and macOS 15.1, initially in English and on a subset of M-series devices. Apple disclosed an on-device foundation model of roughly 3 billion parameters, paired with Private Cloud Compute for heavier requests. Features piled up over the next two years, from Genmoji and Image Playground to writing tools, a rewritten Siri and the Foundation Models framework for developers, and the asset footprint grew with them.
Disk usage is not a new grievance. Siri voice packs,
Liu Gong argues that the interesting number here is not 12GB but what it reveals about the hidden cost of on-device AI.
From a technical standpoint, the root of the dispute is that models have turned from a feature into an asset. When Apple Intelligence launched, the on-device foundation model was about 3 billion parameters, with quantized weights in the low single-digit gigabytes. By macOS 27, the bundled assets had expanded to cover image generation, speech and multilingual resources, compounding to roughly 12GB. Unlike Siri voice packs or language files, these assets are continuously updated, so the footprint only grows. Delete them and the next system update pulls them back, leaving users in a cleanup loop where third-party tools act as a patch rather than a fix.
On the business side, removing the toggle looks less like an oversight than a deliberate product call. Apple Intelligence is now woven into Spotlight, writing tools and Siri, so letting users switch it off piecemeal would create half-broken interfaces. For Apple, on-by-default and invisible protects adoption better than offering a choice. The cost lands on users with the least headroom: Apple Intelligence requires M1 silicon or newer, and the mainstream M1-era machines shipped with 256GB drives, many of them now three years old and permanently near capacity. Twelve gigabytes is close to 5 percent of that sticker capacity, which is not trivial on a machine that has been warning about storage for months.
For the industry, the episode previews a larger fight over who controls on-device AI assets. Phone and PC makers are shipping models into system partitions while rarely saying how much space they take, whether they can be removed, or what breaks if they are. Europe's Digital Markets Act already demands clearer paths for data portability and user choice, and whether local models count as deletable user data is likely the next compliance flashpoint. Liu Gong expects more of these small open-source tools before regulators weigh in, and their real contribution is forcing vendors to answer the question of whether users can switch it off at all.
- The deletion reclaims roughly 12GB, about 4.7 percent of a base 256GB MacBook Air's advertised capacity, or close to 5 percent of its usable ~245GB.
- Compared with macOS 26, where disabling Apple Intelligence in Settings let the system reclaim most model files, macOS 27 keeps the same ~12GB resident indefinitely unless a user takes manual action.
- Against rivals, Microsoft's Phi Silica on Copilot+ PCs is publicly described as a 3-billion-parameter class model quantized into the low single-digit gigabytes, and Google's Gemini Nano on Pixel is similarl
For the industry, the next three months should bring more scripts and tools of this kind, aimed at Windows Copilot+ PCs and Android on-device model directories. Apple's most likely response is technical rather than policy-driven: relocating assets in a future update or restoring a limited toggle in a point release, not opening up deletion rights. Over the following year, managing on-device AI footprint and its removability becomes a system design question, with on-demand model downloads, partition isolation and independent versioning all back on the tabl
Questions worth exploring:
- Where the boundary of user control sits when on-device AI assets become non-removable system storage
- Why Apple removed the toggle: the trade-off between feature penetration and interface consistency
- Whether on-device AI models should count as deletable user data under the EU's Digital Markets Act
This item comes from The Verge AI (published 2026-10-05 13:47:14). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.