Core AI is Apple's on-device ML runtime in iOS 27 / macOS 27 and the successor to Core ML: PyTorch models are exported with Apple's coreai-torch (LLMs: coreai.llm.export) into .aimodel bundles that run on the GPU or the Neural Engine, e.g. Qwen3-8B 4-bit decodes at 94 tok/s on an M4 Max GPU, MLX 90 under the same protocol (apple-silicon-llm-bench, macOS 27 beta 26A5353q, 2026-06-11).
This model has no row on DeviceMark, the on-device LLM leaderboard.
V-JEPA 2 (ViT-L, SSv2 action recognition) — Apple Core AI
V-JEPA 2 (Meta AI) running natively on
the Apple Core AI engine — the zoo's first world model: a self-supervised video encoder that
learns by predicting in representation space (JEPA), here with the Something-Something v2 action
head (174 classes of physical interactions — put/lift/push/roll/cover/pretend…).
- One bundle: ViT-L backbone (3D RoPE attention) + attentive pooler + classifier, ~375M params,
fp16 ~675 MB.
- I/O:
pixel_values_videos [1,16,3,256,256] (16 frames, RGB 0..1, ImageNet mean/std) →
logits [1,174] (labels.json).
- Verified: engine vs PyTorch reference cosine 0.999996, top-5 identical; a synthetic
motion probe (square moving up vs down) flips the predicted direction correctly.
- Speed: ~150–180 ms per 16-frame clip on an M4 Max (GPU) — real-time video understanding.
Use it
One line — this model is the default behind the kit's task op
(import CoreAIOps; no session, no model plumbing, downloads on first use):
let actions = try await CoreAI.recognizeAction(videoAt: videoURL)
Every op, one shape — Cookbook.
▶Run it (source) — the ActionCamera runner
(live camera action recognition, one app for every video model in the catalog):
git clone https://github.com/john-rocky/coreai-kit
open coreai-kit/Examples/ActionCamera/ActionCamera.xcodeproj
# → Run, then pick "V-JEPA 2 ViT-L (SSv2)" in the model picker
# agents / headless (macOS):
cd coreai-kit/Examples/ActionCamera
swift run action-cli --model vjepa2-vitl-ssv2 --video sample.mp4
Build with it — complete; the glue is kit API, copy-paste runs:
import CoreAIKitVision
let recognizer = try await ActionRecognizer(catalog: "vjepa2-vitl-ssv2")
let actions = try await recognizer.classify(videoAt: videoURL, topK: 3)
// actions: ranked [Prediction] — .label ("Pushing [something] from left to right"),
// .probability; 174 SSv2 classes, fully on-device
The take-home is Examples/ActionCamera/Sources/QuickStart.swift
— this exact code as one typed function, no UI; the CLI is an argument shell over it, and
the GUI classifies a rolling 16-frame clip from CameraFeed.
Live camera? Keep the last 16 CameraFeed frames and call classify(frames:) — other
frame counts are uniformly resampled to 16. The bundled sample.mp4 is a synthetic
clip (a hand pushing a block); point --video at real footage for real results.
Integration checklist
- SPM:
https://github.com/john-rocky/coreai-kit → product CoreAIKitVision
- Info.plist:
NSCameraUsageDescription — only for the live camera; the snippet needs none
- Entitlements: none needed
- First run downloads the model — 0.7 GB (Mac) / 0.7 GB (iPhone) — then it loads from the
local cache (Application Support; progress via the
downloadProgress callback)
- Measure in Release — Debug is ~3× slower on per-token host work
Files
| path |
what |
macos/vjepa2_ssv2_fp16.aimodel |
fp16 bundle (macOS / JIT) |
ios/vjepa2_ssv2_fp16.h18p.aimodelc |
iOS AOT bundle (iPhone, A18 Pro+ GPU) |
macos/labels.json, ios/labels.json |
174 SSv2 class names |
macos/metadata.json |
I/O + preprocessing spec |
Live demo app: coreai-video
— camera → live top-3 actions. iPhone 17 Pro: ~0.34 s per 16-frame clip.
Preprocessing
Sample 16 frames uniformly from the clip, resize+center-crop to 256×256, scale to 0..1, normalize
with ImageNet mean [0.485,0.456,0.406] / std [0.229,0.224,0.225], layout [1,16,3,256,256].
Credits
More models in this format: Core AI Model Zoo — 75 models, each with the recipe that produced it.
Want a different model on-device? Open a request — free, open weights only; the export and its measured numbers get published publicly.
