Augmented Mind · Research

Local AI on a Mac, measured

How fast and how good local AI models are on a Mac, measured on real machines: LLM tokens per second as the conversation grows, transcription speed, photo captioning speed and quality, and how much an Apple Silicon Mac changes all of it compared with an Intel one. We run these benchmarks to decide which models our apps should run on the Mac they are installed on, and we publish the numbers because the same questions come up for anyone running AI locally.

The harness, scripts and aggregate results are public at github.com/rememberthis-ai/local-ai-benchmarks. If you have a Mac we haven’t measured, run it and send a pull request with your results.

Posts

  • October 2026
    Generating images, video and voice on a Mac against the best AI APIs — the same prompts through the top hosted image, video and speech models and the best open models on a 64 GB MacBook Pro: every picture, clip and voice memo, what each cost and how long it took.
  • October 2026
    Local LLM speed on a Mac, October 2026 — the same M1 Max runs local models twice as fast as in May, how long a conversation can get before every answer is a wait, a local model driving Claude Code, and photo captioning on Apple Silicon against Intel.
  • May 2026
    Local LLM speed on a Mac: Apple Silicon vs Intel, tokens per second — two weeks on an M1 Max and an Intel i9: ollama against SwiftLM from 4K to 64K tokens of context, Whisper transcription at 33× real time on Apple Silicon, vision models scored for caption speed and quality, and which model fits each amount of memory.