The Verge Launches a Local AI Diary: Hermes, Qwen and a Mac Studio

One author, two computers and an open model on his own hardware. The Verge is opening a series of notes on local AI, in which a writer who has never used AI models on a regular basis tries to work out whether he can do without the cloud. His motivation is simple: he is not willing to hand his most personal data to cloud services. But running powerful models locally is getting easier, and it remains to be seen whether that justifies buying an expensive machine.

The published excerpt has few technical details, and we won't pretend otherwise. The essentials are known: the author calls himself a non-expert, promises to keep notes as he goes, and compares local AI to a journey that is not necessarily worth taking.

What exactly is being tested

The photo caption and the link to the full article point to a working stack. The picture shows an Asus TUF A14 laptop running Hermes Agent, sitting on a child's table among toys. The article's URL mentions Hermes, a Qwen model and a Mac Studio. The setup, judging by all this, is an agent running on top of an open model, with the hardware compared as a "gaming laptop versus desktop Mac."

The text also mentions Apple's marketing argument. A significant part of the pitch for the new desktop Macs is built around how well they handle AI tasks. The excerpt cuts off after that, so we do not cite Apple's figures or any test results here.

• Format: a public diary with interim conclusions, not a one-off review.
• Agent: Hermes Agent, running locally.
• Model: the Qwen family (per the article's URL).
• Hardware: Asus TUF A14 and Mac Studio.
• Key question: is it worth paying extra for a computer with a large amount of RAM?
• Motive: privacy, and reluctance to send personal data to the cloud.

Why the memory question is not rhetorical

Local models have a physical constraint: the weights must fit in memory, and generation speed depends on how fast that memory can deliver data. So a buyer runs into limits of memory capacity and type, not CPU clock speed. The author flatly calls RAM overrated. In 2026 this is a painful subject for anyone building a machine for neural networks.

There is a second difficulty, which the author warns about right in the headline: getting into local AI is at once fascinating, overwhelming and annoying. You have to choose a model, a weight format, the software tooling and an agent. Each decision affects answer quality and speed.

How to use this material

For now it is more a pitch for a series than a finished verdict. But it already yields a sequence of steps for anyone thinking about local AI.

1. Define your task. If you need a private assistant for personal notes and documents, the model requirements will be one thing.

If you need an agent for complex multistep actions, they will be quite different. 2. Start with the hardware you already have.

In this story, a laptop like the Asus TUF A14 plays the role of an entry point, while the Mac Studio is the option for those willing to pay for memory. 3. Don't buy a computer for AI in advance.

First run a small open model and see what you are missing. 4. Follow the rest of the diary.

Measurements on specific configurations will show where running locally can already replace the cloud, and where it can't.

The Verge will publish the full text and test results as the author progresses along this path.