Run AI locally on a Mac for free: we tested Apple's built-in AI on 5 everyday tasks

01Step 1: check that your Mac can run it

Apple's support page lists Mac with M1 or later, the current Mac software (macOS 27), and the Mac and Siri set to the same supported language (English and French are on the list). Apple Intelligence can take up to 14 GB of storage on Macs with M3 or later and at least 12 GB of memory, and up to 8 GB on other supported Macs. Our test machine: a Mac mini with an M4 chip and 24 GB of memory. It came with the software update; we paid nothing extra.

02Step 2: use it without code, in Shortcuts

1. 1
   Open Shortcuts

   Create a new shortcut.

2. 2
   Add Use Model

   Search for the Use Model action, double-click it, then choose On-Device.

3. 3
   Type your request

   In the text field below "Use model". The four-part prompt below works well.

4. 4
   Show the answer

   Add the Show Content action, select Response, then run the shortcut.

Apple describes the On-Device option as handling "simple requests without the need for a network connection". For our test we ran Apple's on-device model through Apple's developer framework (Foundation Models) instead of the Shortcuts app, so we could time each answer. We did not film or test the Shortcuts app itself.

03Our test: 5 tasks, 2 prompt styles, 3 checks

Same fictional inputs as our Claude test: a quiz from study notes, three slides for a family trip, messy notes into a dated to-do list, a confusing building notice, and saying no to hosting a family dinner. Each task once with a lazy prompt and once with a four-part prompt (Context, Goal, Example, Rules). Three checks per answer: right facts, nothing made up, ready to use. One fresh run per prompt, so this is a test, not a study.

Where it slipped: three slides for a family trip

Context: I'm showing my family our weekend trip plan on the TV on Thursday night.
Goal: The text for a 3-slide presentation, one day per slide.
Example: Slide 1 - Friday: Train leaves at 9:10 a.m.
Rules: Use only these facts: Friday train at 9:10 a.m., Saturday museum at 2 p.m., Sunday train home at 5:30 p.m. Add no places, prices, bookings or activities I didn't give you. Three bullets per slide at most.

**Slide 1 - Friday**  
- Train leaves at 9:10 a.m.  
- Travel begins early  
- First day on the move  

**Slide 2 - Saturday**  
- Museum at 2 p.m.  
- Mid-week event  
- Afternoon schedule  

**Slide 3 - Sunday**  
- Train home at 5:30 p.m.  
- End of the trip  
- Return journey starts

The rules said to add nothing and use three bullets per slide at most. It padded every slide to three, and Saturday is not mid-week. With the same prompt, Claude wrote one bullet per day.

Speed on our Mac

The first answer took 7.45 seconds because the model was loading. The next eleven took between 1.07 and 3.01 seconds, about 2.1 seconds on average.

04Bonus: a voice memo to a to-do list, on the Mac

We made a 23.5-second fictional memo with a synthetic voice, added kitchen-style noise, and ran Whisper large-v3-turbo (OpenAI's open-source speech model, MIT license) with the free mlx-whisper package. It transcribed the memo in 2.54 seconds with every detail right (it wrote 1:30 as "1.30"). Asked plainly, Apple's model then dropped the plumber and the 10-to-noon window and added "Now" to two tasks. With this prompt, every detail came back:

Context: This is a voice memo I left myself for this week.
Goal: A to-do list I can check off.
Example: Fri - Dentist at 3:40
Rules: One line per task. Keep every person, day and time I said, word for word. If I gave no day, write "No day". Don't add anything I didn't say.

My voice memo:
[PASTE THE TRANSCRIPT]

Thu - Plumber between 10 and noon, someone has to be home  
Wed - Mia's field trip form is due, she needs a packed lunch that day  
No day - Book the vet for Pepper before November 15th, she's due for her shot  
No day - Pick up milk, eggs, two lemons, and a bag of rice  
No day - Call Aunt Rosa back about Sunday lunch, she said 1.30 works

05Verdict

Worth it for short jobs on your own text: tidying notes, explaining a letter, a polite reply. It is fast, free and already there. It is not for research or long documents: Apple's developer page itself points to Private Cloud Compute or server models when you need more reasoning or a longer context. Use the four-part prompt, and read every list before you trust it.

Want a bigger free chatbot? Google's Gemma 4 12B is about an 8 GB download (Ollama's page), and Meta's Muse Glimmer is under 20 GB once compressed to about 4-bit (Meta's blog). We have not tested those yet.