Skip to main content

Apple Silicon

Apple Silicon Macs (M1, M2, M3, M4, M5) are probably the best platform for this tool. Everything accelerates: video encoding, face detection, and even local LLM inference. Unified memory means no copying data between CPU and GPU.

What you get​

  • VideoToolbox encoding: 5-10x faster than software libx264. Uses the dedicated media engine on the chip.
  • Vision Framework face detection: runs on the Neural Engine, ~10x faster than OpenCV CPU. More accurate too, especially with small or partially occluded faces.
  • Unified memory: no CPU/GPU transfer overhead. Frames stay in the same memory pool whether the CPU, GPU, or Neural Engine is working on them.
  • mlx-vlm for local LLMs: run vision models locally with Metal acceleration for LLM content analysis. No API costs, no data leaving your machine.

Installation​

uv sync --extra mac

The mac extra installs the Apple-specific dependencies (pyobjc bindings for Vision Framework, etc.).

Configuration​

hardware:
  enabled: true
  backend: "apple"   # or "auto" (auto detects it)

auto works perfectly on macOS. It'll find VideoToolbox and Vision Framework without any manual config.

Supported chips​

All Apple Silicon chips are supported:

  • M1, M1 Pro, M1 Max, M1 Ultra
  • M2, M2 Pro, M2 Max, M2 Ultra
  • M3, M3 Pro, M3 Max, M3 Ultra
  • M4, M4 Pro, M4 Max, M4 Ultra
  • M5 and newer

The media engine and Neural Engine get faster with each generation, but even a base M1 runs the full pipeline comfortably.