mlboydaisuke/sam3-CoreAI-official

SAM 3 — official Apple Core AI export

A pre-converted .aimodel segmenter bundle from Apple's official coreai-models export recipe — unmodified, with the exact environment, hashes, and measured performance published.

SAM 3 (Segment Anything Model 3, Meta) is a unified, promptable segmentation model. This bundle is the image + text-prompt variant: give it an image and a phrase (cat, the red car) and it returns instance masks, boxes, and per-instance scores — open-vocabulary, no fixed class list. It runs on the stock Core AI runtime via Apple's CoreAIImageSegmenter (no engine patch).

# from an apple/coreai-models checkout (gated model — accept the SAM license + hf auth login first)
uv run models/sam3/export.py --dtype float16

Why a pre-converted bundle?

An .aimodel is a build artifact, not a pure function of the recipe — the toolchain and OS lowering matter, and the source checkpoint is gated. This is the exact, hash-stamped bundle behind the published numbers, so you can drop it in and reproduce them without a conversion environment or accepting the upstream license just to get the converted graph.

Bundle layout & integrity

A segmenter bundle directory (metadata.json schema 0.2 + the .aimodel + a CLIP tokenizer/):

sam3_float16.aimodel/   main.mlirb + main.hash + metadata.json
tokenizer/              tokenizer.json + tokenizer_config.json
metadata.json           kind: segmenter, assets.main → sam3_float16.aimodel

File	Contents	SHA-256
sam3_float16.aimodel/main.mlirb	float16 weights + graph (~1.5 GB)	be9484b4c5fa75f9011668c481e27de253e1db55fcde39a1ba0cab8698706620
sam3_float16.aimodel/main.hash	graph hash	9f631da3be044ad5de56a1d29af88394e7e725e3cf52f50ac556dcce9cdd568c

Measured

Apple's official image-segmenter CLI, M4 Max, the COCO two-cats image (640×480, resized to 1008×1008 internally), text prompt cat:

Metric	Value
Inference (warm, incl. pre/post-processing)	0.55 s
Output	2 high-confidence instance masks for "cat" (scores 0.97 and 0.96); spurious queries fall below the 0.5 threshold

float16 vs float32 fidelity (same recipe, both exported and run through the engine): top-2 scores differ by ≤ 1e-4 (0.9746 vs 0.9747, 0.9646 vs 0.9645), bounding boxes within 1 px, and the top-3 masks have identical foreground pixel counts. float16 is shipped (half the size, iOS-friendly); float32 reproduces with --dtype float32.

Export environment

• macOS 27.0 beta · Xcode 27.0 beta
• coreai-core 1.0.0b1 · coreai-torch 0.4.0 · transformers 5.x
• apple/coreai-models recipe models/sam3/export.py (unmodified upstream)

Run it

# CLI (from an apple/coreai-models checkout)
swift run -c release image-segmenter \
    --model <downloaded-bundle-dir> --prompt "cat" --image cats.jpg

Or in the CoreAISegment sample app (macOS + iOS): pick an image, type a prompt, tap Segment.

import CoreAIImageSegmenter
let segmenter = try await ImageSegmenter(resourcesAt: "<bundle-dir>")
let result = try await segmenter.segment(image: cgImage, prompt: "cat")
// result.segments: [Segment] — each has .mask, .box, .score

iOS bundles should be AOT-compiled before device use: xcrun coreai-build compile <ir>.aimodel --platform iOS --architecture h18p (h18p = iPhone 17 Pro), then point metadata.json assets.main at the .aimodelc.

License

SAM 3 is distributed under Meta's SAM License (see LICENSE and the upstream model card). This bundle is a format conversion of those weights and inherits that license; review it before use.

---

Maintained alongside coreai-model-zoo (see official/ for the other unmodified official-recipe conversions).