Author Tests 125B-Parameter Qwen on $12,000 Mac Studio for Private Local AI
Updated
Updated · The Verge · Oct 11
Author Tests 125B-Parameter Qwen on $12,000 Mac Studio for Private Local AI
2 articles · Updated · The Verge · Oct 11
Summary
A 125-billion-parameter Qwen 3.8 Flash Next model, about 105GB in size, was installed through the self-hosted Hermes Agent on an M5 Ultra Mac Studio with 256GB of unified memory.
Privacy drove the setup: running models locally let the author analyze sensitive financial records and embargoed laptop specs without sending data to OpenAI, Google, Microsoft, Anthropic or other cloud services.
Hermes handled a few early tasks, including a 7:30 a.m. email-calendar-weather briefing and reorganizing a 400-plus-game Steam library by genre after temporary access via a Steam web API key.
The experiment also exposed limits: the daily briefing initially failed because macOS was asleep, later broke again, and a larger effort to automate repeated laptop benchmark tests remains unfinished.
The broader test is whether expensive high-RAM machines such as Apple’s AI-focused desktops can make local agents useful enough to justify their cost for privacy-conscious users.