Local AI on the M6 Mac Mini: Coding, Images & More With 24GB

Summary

The video tests local AI workflows on a Mac mini described in the subtitles as having an M6 chip and 24 GB of RAM. The creator compares model sizes and settings across writing, document analysis, image generation, transcription, and text-to-speech.

Writing and coding

In LM Studio, an 8B model generated a usable first draft quickly, running at about 28 tokens per second. On an M4 Mac mini with 16 GB of RAM, the same test ran at 18 tokens per second—about 56% slower. However, the model sometimes gave inaccurate answers, including incorrect calculations and code-related results. The creator attributed those errors to the model rather than the computer.

Document analysis

An 8B model handled a sample business report without performance issues, though its accuracy was imperfect. A 14B model could search longer documents, but response times increased from about 7 seconds for the shortest text to nearly 30 seconds for the longest. The longest task used around 20 GB of memory, leaving little room for other apps.

A test with a 27B model froze the Mac and required a restart. The results suggest that 8B and 14B models are more practical on this 24 GB system.

Image generation

Draw Things generated a 1024 × 1024 image with SDXL in under 20 seconds, using about 17–18 GB of memory. Increasing the resolution to 2K raised memory use to roughly 22 GB, with about 2 GB of swap.

The creator preferred the more realistic results from Flux 2, a 4B model, which generated images at a similar speed.

Audio

MacWhisper transcribed a video just over seven minutes long in under 12 seconds, using about 9 GB of memory. The creator reported that the transcription was fully accurate.

For voice generation, Kokoro and Supertonic produced audio from a 200-word script in about 13 seconds. A script of more than 2,000 words took just over six minutes and used about 10 GB of memory.

Overall assessment

The Mac mini can handle useful local AI tasks when model sizes and settings are chosen carefully. The creator said local tools had replaced some cloud services for image creation and voiceovers, offering offline use and greater privacy.

Cloud AI remains preferable for communication, writing, and research because it is more capable and connected to the internet. Memory pressure is the main limitation, especially when running larger models or multitasking.

Practical guidance

1. Compare model performance across hardware configurations, but evaluate accuracy as well as speed.
2. Use 14B models cautiously for longer documents; expect slower responses and high memory use.
3. Avoid the tested 27B model on this 24 GB system, as it froze the computer.
4. Start image generation at manageable resolutions. The creator preferred Flux 2 for image quality.
5. Try MacWhisper for fast local transcription and Kokoro or Supertonic for voice generation.

Main speaker/source: The video’s presenter, identified in the metadata as Tommy Doan.