The promise of local AI is straightforward: use capable models without sending private information to the companies operating cloud services. A hands-on report from The Verge shows that the reality is both more useful and more demanding. Running an agent on a powerful desktop made several sensitive tasks possible, but it also required expensive hardware, technical setup, machine permissions and repeated troubleshooting.

The experiment began on an M5 Ultra Mac Studio with 256GB of unified memory, part of a broader test that also includes other Mac and Windows systems. The Verge notes that Apple promotes its newest desktops for local AI, while Windows machines with as much as 128GB of RAM are arriving for agentic workloads. The computer used in the test cost about $12,000, though the report explicitly says a machine that expensive is not required for every task.

A massive AI model block is fitted into a high-memory desktop among many competing model choices.
The Verge tested a roughly 105GB, 125-billion-parameter model on a Mac Studio with 256GB of unified memory.

The author installed Hermes Agent, an open-source, self-hosted desktop application available for macOS, Windows and Linux. Hermes can be used without a software fee when paired with local models. On the high-memory Mac Studio, the test started with Qwen 3.8 Flash Next, described in the report as a 125-billion-parameter model occupying roughly 105GB. The abundance of model choices quickly became part of the difficulty.

Basic setup was relatively quick, and the agent could be controlled from a phone through a Telegram bot. Choosing what to do with it was harder. The first project was a scheduled morning briefing that reviewed email and calendar information and added weather. It repeatedly failed until the author realized the Mac could not be asleep when the 7:30 a.m. task was supposed to run.

A more successful job involved a Steam library containing more than 400 games. The agent proposed ways to organize the collection, then sorted titles by genre while preserving custom groups such as favorites, cooperative games and games played with the author’s wife. Completing the work required a Steam web API key, which the author revoked after the one-time task was finished.

Sensitive records remain inside a locked computer as a temporary access key is withdrawn.
Local processing enabled work with financial and embargoed data, while the Steam task showed the value of revoking access after use.

Local processing made the biggest difference when the data itself was sensitive. The Verge’s author used the agent to analyze financial records and create a laptop specification spreadsheet containing information under embargo. Those were tasks the author said would not have been entrusted to a cloud service. Keeping the material on the computer changed the decision from avoiding AI entirely to using it for practical work.

That privacy advantage does not remove the need for access controls. An agent that reorganizes a game library or reads email must receive permission to reach those resources. The report’s Steam example demonstrates a limited approach—grant access, complete the task and revoke the key—but it does not claim that every local tool automatically follows that pattern or that local execution eliminates security risk.

The most ambitious project remains unfinished: automating a set of laptop benchmark tests that are normally run manually several times to calculate averages. Walking the agent through the procedures and producing Python scripts has taken substantial effort. The Verge’s early conclusion is therefore measured. Local AI can already handle useful private work, but even a large model on premium hardware is still a tool that breaks, needs supervision and falls well short of a universal assistant.