← All articles

Mikhail Polyakov

An AI assistant on your own computer instead of the cloud: what worked for a reviewer at The Verge and how to work out whether you need it

Level: BeginnerReading time: 8 min

Topics: AI, Automation, Small business, Security

A silver cube-shaped desktop computer on a wooden desk, next to a folder of blank sheets that a warm light stretches toward from the computer, and a mug; above the computer a crossed-out glowing cloud is dissolving

Short answer: The Verge reviewer Antonio Di Benedetto, who did not want to hand his personal data to cloud AI services, installed the open-source Hermes AI agent with a Qwen model directly on his own computer, a Mac Studio with 256 GB of memory. Getting it all running was quick. The harder part was figuring out what to give it to do. Of what he described, the useful parts turned out to be sorting a large game library and working with data he would not have uploaded to the cloud. The morning briefing was of limited use and broke several times, and the workflow for his laptop tests is not finished yet. Below is what exactly he did, what access he had to give the agent, and how to work out whether a computer like this would pay off for you.

What local AI means in this story

A chatbot usually runs on the servers of the company that made it: you type a question, it goes to that company, and the answer comes back. A local model runs on your computer, and the model itself does not send your questions to someone else's server. But the agent around it may be connected to outside services: in the author's case, for example, it was controlled through a Telegram bot and worked with his email and Steam. Anything that passes through such connections goes outside just as before.

An AI agent here means a program that does not just answer questions but also carries out actions on the computer by itself: it opens programs, reads files, runs scheduled jobs. Hermes is that kind of agent. According to The Verge, it works on macOS, Windows and Linux and is free if you only use local models.

What the author installed

All the facts below come from The Verge's article of October 11, 2026. This is one person's diary of his first weeks, not a test and not a study.

  • Who. A laptop reviewer who spent more than 15 years in the photography industry before The Verge. He says himself that he is not an AI expert.
  • Computer. A Mac Studio with an M5 Ultra processor and 256 GB of unified memory. The author himself calls it a $12,000 computer.
  • Model. Qwen 3.8 Flash Next: 125 billion parameters, around 105 GB. Parameters are the model's internal settings; according to the author, models with more of them can generally do more.
  • Control. Getting Hermes running with Qwen was quick, and he could even control the agent from his phone through a Telegram bot.

Then, in the author's words, the difficulties began. There are so many models that the choice quickly becomes overwhelming. And the main question came up in front of an empty text box: "What do I do with this thing?"

What he gave the agent to do and how it went

Task What happened What got in the way
Morning briefing: email, calendar, weather Started working once the author found the first cause of failure, but then broke several more times At first it would not generate: the computer was asleep at 7:30, when the job was scheduled to run. The author himself writes that the briefing is not all that useful
Sort more than 400 Steam games by genre The agent suggested options, the author chose genres while keeping his own categories, the job took a few minutes It needed permissions on the computer and a Steam access key
Crunch numbers from financial records, compare laptop specs Done, the author gives no details No difficulties described. This is exactly the data the author would not give to the cloud
Workflow for running laptop tests Not ready Explaining the test procedure to the agent and getting Python scripts out of it turned out to be hard

About the financial records and the laptop data, which were under embargo until a set date, the author said plainly: the fact that they stay on his computer was the difference between using the tool or not.

His overall conclusion: even powerful local models do not turn the agent into a magical assistant that can do anything, and things still do not always work as intended. He does not intend to have the agent write or edit texts for him.

What the hardware costs

Prices from the same article:

  • the author's computer, in his words, costs $12,000;
  • at an Apple demonstration the author attended, Apple showed a cluster of four Mac Studios for local AI, which is nearly $50,000 worth of hardware;
  • according to The Verge, Windows machines in the RTX Spark line with up to 128 GB of memory, aimed at AI agents, are coming soon.

The author separately plans to try smaller models on other computers, from a Mac Mini to a gaming laptop. How well they will handle the same tasks, the article does not say yet.

How to estimate the payback for yourself

This is my calculation template, not data from the article. It has no numbers in it: put in your own.

Hours saved are not money by themselves. Only count the benefit the company can confirm with invoices and reports:

  • B: confirmed benefit per month. This includes expenses that were avoided (overtime, payments to a contractor, a person you did not have to hire) and extra income from work done in the freed-up time, minus the cost of that work. If the freed-up time went to nothing, B is zero, however many hours the agent saved.
  • S: support per month: hours spent on setup and fixing, multiplied by the rate of whoever does the fixing. The author's briefing broke several times, so this is not zero.
  • R: other monthly costs: electricity, maintenance, paid connections.
  • C: the price of the AI computer including setup.

Net benefit per month: B − S − R. Payback period in months: C divided by the net benefit. It only makes sense to calculate it if the net benefit is above zero; if it is zero or less, the computer will not pay off.

If the data can be given to a cloud service, also compare with a subscription for the same tasks: work out its price over the same period together with the hours spent on setup.

What to do before buying

These are my recommendations.

  1. Task first, computer second. Write down three to five tasks you would not give to the cloud because of the data. For the author, these were exactly the ones that made the difference.
    • Who does it: the manager together with whoever currently does this work.
    • How to check: for each task, it is written down how many hours a month it takes now.
  2. Try it on what you already have. Hermes is free with local models, and the author himself plans to try smaller models on other computers too. A trial run on an existing computer will show whether the model can handle your task before you spend money on a new one.
    • Who does it: an employee who knows their way around computers, or an outside specialist.
    • How to check: ten real cases from your own work, with a note for each on whether the agent got it right.
  3. Check what goes outside. Before working with confidential data, find out which outside services the agent is connected to (messenger, email, cloud models) and what it sends to them.
    • Who does it: whoever is responsible for computers and accounts, or an outside specialist.
    • How to check: the list of connections in the agent's settings is written down; anonymized data is used for the trial, and real data only after the list has been checked.
  4. Grant access temporarily. The author's agent needed permissions on the computer and a Steam key, and he revoked the key right after the job. The same goes for work systems: access only for a specific task, then revoke it.
    • Who does it: whoever is responsible for accounts in the company.
    • How to check: after the task, the agent's key shows as revoked in the service's settings.
  5. Allow time for breakdowns. In the author's case, the scheduled job did not run while the computer was asleep; how your computer behaves depends on its settings. Decide in advance who checks that the briefing or report has arrived.
    • How to check: for a week, write down on which days the result did not arrive.

When this does not concern you

If your data can safely be given to a cloud service and you do not have many tasks, a cloud subscription may turn out simpler: no need to buy a computer and fix the jobs yourself. A computer of your own for AI is worth considering when the data must not leave, as with the author's financial records and embargoed data. But even then only after checking what information the agent passes to connected services and what access it has been given: a computer on your desk does not by itself keep data inside.

Summary

A reviewer at The Verge installed the Hermes AI agent with a Qwen model on his own computer and described his first weeks. Getting it all running turned out to be easy, but finding useful tasks took more work: what made the difference were tasks with data he would not give to the cloud. The briefing broke, and the workflow for laptop tests is not ready yet. Before buying a computer for AI, write down your tasks, count the hours and try it on what you already have.

I work on AI agents and automation. If you would like to see my projects or discuss your own task, take a look at my portfolio.

Sources