What changed
2025, iOS 26. Apple's Foundation Models framework gave apps access to one model: Apple's own on-device model, about 3 billion parameters, the one behind Apple Intelligence 1.
June 2026, WWDC. Apple announced Core AI, described as the successor to Core ML 5. Developers convert a model, or use Apple's export recipes, ship it as .aimodel files, and run it on the CPU, GPU and Neural Engine through one Swift API 23. Apple's coreai-models repository has recipes for popular open models 3:
- Language: Gemma 3, 3n and 4; Qwen 2, 3 and 3.5; Mistral; Mixtral; Phi; gpt-oss; OLMo 2; SmolLM2.
- Speech: Whisper and Parakeet.
- Vision and images: CLIP, SAM 3 and FLUX.2.
14 September 2026. iOS 27 shipped 4. Core AI needs iOS or macOS 27 and Xcode 27 3.
It is not a button that lets you import a model. The app's developer chooses the model, converts it, and ships it with the app or downloads it later. Open-source apps such as PocketPal AI already let users load their own model files, and did before iOS 27 6.
What it does for privacy
When the model runs on the device, the prompt, any document you hand it, and the answer never have to travel to a server to be processed. No server receives or logs the raw text, and once the model is on the phone the feature can work without a connection. For sensitive material such as health notes, legal documents or a journal, that removes the largest exposure: a third party receiving the raw text at all.
Three things it doesn't settle
1. The app can still send data elsewhere. On-device inference says nothing about analytics, crash reports, cloud sync or the app's own backend. Two quick checks: the App Store privacy label, and whether the feature still works in airplane mode.
2. Not every AI feature on the phone is local. Apple's own new Siri features in iOS 27 run both on the device and in the cloud, on Apple's Private Cloud Compute 4. An app running its own bundled model through Core AI takes a different path. Don't assume one from the other.
3. Memory decides what fits, and small models make more mistakes. The weights have to fit in the phone's memory alongside everything else:
Memory for a model's weights at 4-bit, by size
Weights only. The conversation cache and the app itself need more on top.
Source: Arbiter arithmetic: parameters × 4 bits ÷ 8 bits per byte
Show the data
| label | value | display | params_billion | bits |
|---|---|---|---|---|
| 1B parameters | 0.5 | 0.5 GB | 1 | 4 |
| 3B parameters | 1.5 | 1.5 GB | 3 | 4 |
| 4B parameters | 2.0 | 2 GB | 4 | 4 |
| 8B parameters | 4.0 | 4 GB | 8 | 4 |
| 14B parameters | 7.0 | 7 GB | 14 | 4 |
| 27B parameters | 13.5 | 13.5 GB | 27 | 4 |
| 70B parameters | 35.0 | 35 GB | 70 | 4 |
Phones run small models, and local does not mean correct. In our own measurements, a local 7B judge scored identical facts 0.0 or 1.0 depending on where a citation sat 7.
Why this matters for health data
Keeping inference on the device cuts down where protected health information travels. It does not make an app HIPAA compliant on its own: storage, sync, access control and audit still apply. Arbiter's agent follows the same idea on a server. Every model call goes to a local model, and the agent refuses to start if a model endpoint is anything other than local or covered by a business associate agreement (HIPAA-ENDPOINTS). Not legal advice.
Show your work
How this was done
What Core AI is and supports comes from Apple's own documentation and its public apple/coreai-models repository, read on 2 November 2026 23. The ship date is from press coverage of Apple's announcement 4. The memory figures are arithmetic, not measurements: parameters × 4 bits ÷ 8 bits per byte, in decimal gigabytes. Real quantized files are a little larger because some layers stay at higher precision. The cache that grows with the conversation is extra. Apple does not publish iPhone memory sizes, so the chart doesn't draw one.
What this does not tell you
We did not benchmark Core AI's speed or accuracy on a device. The list of supported models is the repository's at the time of reading and will change. Whether a given app runs its model locally can only be checked app by app.
Data
Every number in the charts, as CSV:
Agent policy
This read is evidence for these rules in the agent's policy:
Sources
- Apple Machine Learning Research, "Updates to Apple's On-Device and Server Foundation Language Models" (9 Jun 2025)
- Apple Developer, Core AI documentation
- Apple, coreai-models repository (export recipes and model catalog)
- 9to5Mac, "Apple confirms iOS 27 release date: September 14" (9 Sep 2026)
- InfoQ, "Apple Launches Core AI for Apple-Silicon Optimized On-Device Generative AI" (June 2026)
- PocketPal AI (open source)
- Arbiter post-mortem: the 7B judge scored identical facts 0.0 or 1.0 depending on citation placement
Corrections
None so far. Found an error? Tell me and it will be corrected here, with the date.