Hermes Agent, Qwen 3.8 Flash Next tested locally on M5 Ultra Mac Studio

The experiment sits at the intersection of two hardware pitches: Apple's claim that its new Mac desktops are good for local AI, and an incoming line of RTX Spark Windows machines with up to 128GB of RAM aimed at agentic AI. The writer's M5 Ultra Mac Studio has 256GB of unified memory, which let him run the 125-billion parameter Qwen 3.8 Flash Next model at around 105GB in size, a scale that smaller machines in his planned tests — an M6 Mac Mini, an M5 MacBook Air, and an Asus TUF Gaming A14 with AMD Strix Halo processor — may not match.

He frames local AI as a privacy trade-off rather than a capability leap. The tasks he values, crunching financial records and building a spec spreadsheet for an embargoed laptop, are ones he says he cannot or will not trust to a cloud service. Yet even the large local model did not behave like a magical assistant: his daily briefing broke multiple times, and the benchmark automation script remains a work in progress.

He also draws a line at how he interacts with it, treating Hermes strictly as a tool — no pleasantries, no digital friendship — and says he hopes to remain vigilant and cautious as he continues testing. What he has not yet resolved is whether the usefulness justifies spending heavily on high-RAM hardware.