Testing Local AI on a Mac Studio With Hermes and Qwen
The Verge published a testing diary entry on Oct. 11, 2026 about making local AI a daily tool, starting with setup. The Verge
The work centers on an M5 Ultra Mac Studio, with added tests on other Mac and Windows systems. The main Mac Studio has 256GB of unified memory, shared by processing and graphics.
The writer installed the open-source Hermes Agent on the Mac Studio. Hermes Agent is a self-hosted AI agent desktop app for macOS, Windows and Linux. It is free to use with local LLMs, AI models that run on your own hardware.
Model choice was handled through Hermes' onboarding interface and model picker. The pick was Qwen 3.8 Flash Next, described as a 125-billion-parameter model at about 105GB. Parameters are the tuned values that shape what a model can do. The Verge
The writer then made Hermes and Qwen controllable from a phone using a Telegram bot.
The broader context here is memory fit. A 105GB model on a 256GB system leaves room for the operating system, the agent software, and context, the working memory for the current task. Local systems must share memory with everything else running, so that headroom counts.
In my view, the tooling cuts friction more than raw size does. The onboarding and picker replace manual download and setup. Free use with local models allows repeated tries without licensing steps, so builders can test and discard quickly.
Looking at what this means for agent workflows, self-hosted and cross-platform support changes where testing happens. The same Hermes desktop runs on macOS, Windows and Linux. That helps people with mixed machines, keeps testing on daily computers, and simplifies comparison.
In my view, phone control through Telegram is the detail to watch. A fixed workstation can be tasked from elsewhere. Control leaves the desk while the machine stays put. The work then becomes operational: queuing requests, handling errors, limiting autonomy, and reviewing results. Remote access raises the need for logging and oversight.
Looking at what this means for learning local AI, step-by-step notes have practical value. Model choice, memory budget, software, and remote access interact, and small defaults add up. A public log lets others compare notes and avoid repeat setup. If routine use follows, capable local systems become easier to reach for each day.
