Skip to Content
Platform.Ontology

On-device ML runtime (Core ML, LiteRT)

Runs trained models on the phone's neural hardware.

Works without askingOn-device AIML infrastructureAndroid and iPhoneWhile you use it

In one line

Can run trained AI models on your phone's own chip

Runs trained models on the phone's neural hardware.

Developers ship models with the app and run them on CPU, GPU or the neural chip. NNAPI is deprecated in favour of LiteRT, which targets vendor accelerators.

What it can and can't do

What the user does: Nothing. It happens inside the app when you use a smart feature

It can

  • Run models for photos, speech or text using the phone's neural hardware
  • Work offline
  • Use faster chips on some Android phones, like Qualcomm and MediaTek

It cannot

  • Use the Neural Engine on iPhone by choice; the system handles it automatically
  • Run without a model the app brought along
  • Send your data to a server just because it uses this tool

How it differs between platforms: iPhone uses Core ML, which picks the Neural Engine automatically. Android uses LiteRT, with extra support for some chip makers.

On Android and iPhone

AndroidiPhone
StatusSupportedSupported
First versionAndroid 8.1 (API 27)iOS 11.0
Permission or entitlementNoneNone
Protection levelnonenone
How access is grantednonenone
What the user seesNoneNone
How long it lastsn/an/a
How to take it backn/an/a
Store policyNone listedNone listed
APILiteRT (formerly TensorFlow Lite), LiteRT NPU delegate, NNAPI (deprecated in Android 15), MediaPipeCore ML, MLX (macOS), Core ML Tools, BNNS Graph
NotesLiteRT Next adds vendor NPU accelerators (Qualcomm, MediaTek).Core ML uses the Neural Engine automatically.

Data you get

What comes back: Model output
Shape: object
How often: per run

FieldTypeUnitRangePrecision
tensorarrayfloatn/an/a

Technical limits

  • Model size and memory limits

The story

What it is

The built in engine that runs trained AI models on the phone's chips.

For you

You never see it. Face unlock, photo search and live camera effects run quickly and privately because a model is running on your phone.

For makers

Makers ship a model inside the app and run it fast on special chips. This keeps data local and avoids server costs, though a model in the app can be copied.

Where you see it

TypeExampleApps
gamesA game runs a smart opponent or upscaling model on the phone.None named
creativeA photo app removes a background in real time.Lightroom, Snapseed
healthA fitness app counts reps from the camera with no upload.None named
securityA banking app checks for face spoofing on device.None named
musicA music app recognizes a song or splits vocals.Shazam
ar-vrAn AR app tracks your hands and face live.Snapchat
adsAn ad model on the phone ranks which ads to show without sending data out.None named

Risks

  • Model theft from the app
  • Hidden to the person

Android and iPhone in short

  • Android: LiteRT replaces NNAPI and adds chips from Qualcomm and MediaTek.
  • iPhone: Core ML uses the Neural Engine automatically.

What the platform says

  • Apple: Core ML. Apple describes Core ML as the way to run machine learning models on device with the Neural Engine.
  • Google: LiteRT. Google positions LiteRT as its on device runtime for fast, private machine learning.

Good uses and abuse

Legitimate uses

  • Vision
  • Audio
  • Recommendation

Known abuse

  • Model theft from the app bundle

Scores

ScoreValueMeaning
Sensitivity1 of 5How revealing the data is
Utility5 of 5How useful it is to honest apps
Friction0 of 5How much the user has to do
Fingerprint bits0Rough identifying entropy, when known

History

  • 2017-06, iPhone iOS 11: Core ML
  • 2024-09, Android 15: NNAPI deprecated
  • 2024-11, Android LiteRT: TFLite renamed to LiteRT

Sources

How well checked: Based on official documentation.

On this page