Who is behind this site?
AI on Mac is run by Julian Dominic Altmann / Altmann Digital Studio. I operate this site as an independent technical blog about local AI on Apple silicon Macs. I research models, tools, and workflows, check vendor claims and benchmarks against primary sources, and work up the results with AI agents so that Mac users, developers, and privacy-focused readers can actually use them. The process is described under How this site is made.
My focus is on local LLMs, speech AI, vision models, and the tools that make them work: Ollama, LM Studio, MLX, llama.cpp, and the Apple silicon platform.
Who is this blog for?
Mac users
You want to know what actually runs on an M1, M2, M3, M4, or newer Apple silicon Mac.
Privacy-conscious readers
You want to use AI while keeping as much data as possible on your own machine — with as little cloud dependency as possible in the workflows being tested.
Developers
You're looking for setups, APIs, local servers, benchmarks, model comparisons, and macOS integration tips.
Creators and power users
You want to use local AI for writing, coding, image analysis, transcription, automation, or personal assistants.
Why this blog?
Many AI workflows today rely on cloud services such as ChatGPT, Claude, or Gemini. These services are powerful and convenient, but they often involve external processing, recurring costs, and less control over data, models, and infrastructure.
With Apple silicon, running AI locally on a Mac has become practical for many users. Unified memory, efficient GPUs, and tools such as Ollama, LM Studio, MLX, and llama.cpp make it possible to run many models directly on macOS.
This blog separates theoretical memory estimates, vendor data, reproducible setup steps and measurements actually performed. It helps readers check what will run on their own Mac, which memory limits matter, and when a cloud service makes more sense.
What you'll find here
- Setup guides — Step-by-step instructions from installation to stable day-to-day use
- Model comparisons — Vendor and independent benchmarks, transparent cost calculations and clearly explained evidence limits
- Hardware guides — Mac mini, MacBook Pro and Mac Studio evaluated using documented specifications, memory calculations and explicitly labeled measurements
- Tool explanations — Clear explanations of Ollama, LM Studio, MLX, llama.cpp, and other local AI tools
- Workflow examples — Real use cases for coding, writing, research, automation, and knowledge work
- Cloud vs. local comparisons — Honest assessment of when cloud is better and when local inference is actually worth it
How I check and test
Original measurements: A figure described as a Mac test needs the chip, unified memory, model ID and quantization, runtime version, context, prompt/test file, number of runs and result. A tokens-per-second claim without that information cannot be reproduced.
Calculated estimates: Model weights, KV cache and operating-system headroom are model-dependent estimates. The RAM estimator is a planning aid, not evidence of load time or decode throughput.
External numbers: Vendor benchmarks and provider prices should be attributed with their source and date. Results from different test sets or regions are not combined into a fabricated universal ranking.
Corrections and evidence limits
When pricing, model IDs, licensing or API limits change, the particular claim needs fresh verification against its primary source. A successful automated build shows that the website can compile, not that each published technical statement is accurate.
To suggest a verifiable correction, send the article URL, disputed claim and a supporting source to julian@altmann-digital-studio.de. Corrections are evaluated in context; a modified date alone is not a certificate of quality.
How this site is made
AI agents help research, draft, translate and illustrate the articles. Automated checks cover file references, the build and certain citation formats. Verifying whether a specific source actually supports a claim is a separate editorial task; a passing build does not substitute for individual review. Price comparisons also require a date, region and dependable primary sources.
I set the topics, rules and quality checks, correct mistakes, and carry editorial responsibility for everything published here. That is why every article states that it was researched and written with AI assistance. My own measurements are marked as such in the article; vendor claims, API data and community figures are not presented as my own tests.
Exactly which content is made with AI and how the EU AI Act disclosure is implemented is described on the AI transparency page.
The goal is not to produce as much content as possible, but to provide useful guidance for local AI on the Mac. If you spot an error, corrections are welcome.
Tech stack
Astro
A static site generator for a fast, lightweight website — little JavaScript by default.
Cloudflare Pages
Hosting and CDN for fast delivery of the static site.
Ollama
A tool that simplifies running large language models locally on your own machine.
LM Studio
A desktop app for loading, testing, and serving local AI models with a user interface and local server mode.
MLX
Apple's machine learning framework for Apple silicon, optimized for unified memory and efficient execution on Apple platforms.
llama.cpp
A major open-source foundation for efficient LLM inference with minimal setup across many hardware platforms.
Privacy: local AI vs. running this website
When I write about local AI, the goal is usually to keep models and data on your own Mac. That is separate from how this website is operated.
The website itself may use optional services such as Google Analytics or future Google advertising to measure reach and support the project financially. These services are only loaded after your consent. More details are available in the privacy policy.
Funding and affiliate links
This site may contain ads and affiliate links. If you buy something through such a link, I may earn a commission at no extra cost to you. Recommendations should still remain editorially justified. A positive review cannot be bought. Sponsored content would be clearly labeled as such.
Affiliate links or ads do not change the editorial assessment. If a tool has weaknesses, they are still mentioned.
Contact
Have feedback, a tool recommendation, or an idea for a test?
Write an email to: julian@altmann-digital-studio.de
Legal and transparency
For corrections, use the contact page. Operator details are in the imprint, and the AI disclosure explains how articles are produced. Running a model locally is a separate privacy question from visiting this website.
- Legal notice — Operator and legal information
- Privacy policy — Website privacy and consent
- All English articles — Guides, local models, comparisons and hardware