Running SAM 2.1 on Mac — Apple Silicon GPU Acceleration and Meta SAM 3 Comparison

Wed, 08 Apr 2026 00:00:00 +0900

Overview

Meta’s Segment Anything Model (SAM) changed the game for image segmentation. SAM 2.1 can be run locally on your own machine, while the latest SAM 3 is available through Meta’s online playground. In this post, I run SAM 2.1 on an Apple Silicon Mac with MPS GPU acceleration and compare it with the SAM 3 online demo.

SAM 2.1 Local vs SAM 3 Online — Architecture Comparison

flowchart LR
 subgraph Local["SAM 2.1 Local"]
 A["User Input <br/> Points/Boxes"] --> B["SAM 2.1 Tiny <br/> 74.5 MB Model"]
 B --> C["PyTorch MPS <br/> Apple Silicon GPU"]
 C --> D["Gradio Web UI <br/> localhost:7860"]
 end

 subgraph Cloud["SAM 3 Online"]
 E["User Input <br/> Text/Click"] --> F["SAM 3 <br/> Meta Server"]
 F --> G["Cloud GPU <br/> Inference"]
 G --> H["Web Browser <br/> aidemos.meta.com"]
 end

 style Local fill:#e8f5e9,stroke:#2e7d32
 style Cloud fill:#e3f2fd,stroke:#1565c0

SAM 2.1 on Apple Silicon Mac

The ice-ice-bear/sam2-mac-test repository provides a ready-to-run SAM 2.1 setup for Apple Silicon Macs.

Key Features

MPS GPU Acceleration: Uses PyTorch’s Metal Performance Shaders backend to run inference on M1/M2/M3/M4 GPUs
Multi-point Segmentation: Place include/exclude points for fine-grained segmentation with undo/clear support
Segment Everything Mode: Segment all objects in an image at once
Gradio Web UI: Browser-based interface accessible right away
SAM 2.1 Tiny Model: Lightweight 74.5 MB model, auto-downloaded on first run

Quick Start

git clone https://github.com/ice-ice-bear/sam2-mac-test.git
cd sam2-mac-test
uv sync
uv run python app.py

Open http://127.0.0.1:7860 in your browser to access the Gradio UI.

Performance

Benchmarks on an M1 MacBook:

Task	Time
Single point segmentation	~1.6s
Multi-point update	~1.5s per update

The Tiny model keeps memory usage low, and MPS acceleration provides significant speedup over CPU-only inference.

Tech Stack

SAM 2.1: Via the Ultralytics library
PyTorch MPS: Apple Silicon GPU backend
Gradio: Web UI framework
uv: Package manager

Meta SAM 3 Online Playground

Meta offers the latest SAM 3 as an online demo at aidemos.meta.com/segment-anything.

What Sets SAM 3 Apart

Text-prompt Segmentation: Find objects using natural language — “find animal”, “find person”
One-click Effects: Apply blur, clone, desaturate, and more with a single click
Motion Trails: Add motion effects to segmented objects
Contour Lines / Bounding Boxes: Various visualization options
Video Segmentation: Track Anything feature for object tracking in video
Community Templates: Use effects created by other users

SAM 2.1 Local vs SAM 3 Online Comparison

Aspect	SAM 2.1 Local	SAM 3 Online
Environment	Local Mac (Apple Silicon)	Meta cloud servers
GPU	MPS (M1/M2/M3/M4)	Cloud GPU
Model Size	Tiny 74.5 MB	Full-size (undisclosed)
Input Methods	Point click, box	Text, click, box
Text Prompts	Not supported	Supported
Post-processing Effects	None	Blur, clone, desaturate, etc.
Video Support	Not supported	Supported
Privacy	Data stays local	Uploaded to Meta servers
Internet Required	Only for model download	Always
Customization	Full code access	Limited

Which One Should You Choose?

SAM 2.1 local is the right choice when:

You don’t want sensitive images leaving your machine
You need to integrate segmentation into an automated pipeline
You want to modify or extend the model
You need to work offline

SAM 3 online demo is the right choice when:

You want to find objects using text prompts
You need quick access to effects like blur and cloning
You need video segmentation
You want to try it out without any installation

Wrapping Up

Running SAM 2.1 locally is a practical option for Apple Silicon Mac users. The 74.5 MB Tiny model delivers usable segmentation results, and MPS acceleration makes good use of the GPU. The SAM 3 online demo takes it further with text prompts and a rich set of effects. Depending on your use case, combining local and cloud approaches gives you the best of both worlds.

Pytorch-Mps on ICE-ICE-BEAR-BLOG