Science & Tech · AI · 18 hrs ago
Testing local AI brings promise, complexity and frustration
A technology writer has begun testing AI models that run on personal computers instead of relying on cloud services.
The writer is using an M5 Ultra Mac Studio, which has 256GB of unified memory, and also plans to try other Mac and Windows devices.
The first setup uses Hermes Agent, a free, open-source app, with a 125-billion-parameter Qwen model that takes about 105GB of space.
Running AI locally can help keep personal data on the user’s device, but choosing among many models and figuring out what to ask them can feel overwhelming.
The writer has set up a daily briefing delivered through Telegram as an initial task.
The tests are meant to explore whether local AI is useful enough to justify the cost of computers with large amounts of memory.
Further experiments with smaller models and other devices are planned.
The Verge writer is testing local artificial intelligence models on Mac and Windows computers.
The writer says running models locally could offer more privacy than relying on cloud services.
The experiment began with the open-source Hermes Agent and a 125-billion-parameter Qwen model on an M5 Ultra Mac Studio.
The writer found the range of available models overwhelming and was unsure what tasks to give the system.
The writer set up a daily morning briefing and made Hermes controllable through a Telegram bot.
- Who
- A writer for The Verge, whose name is not stated in the article.
- What
- The writer is experimenting with locally run AI models and an AI agent.
- When
- The Verge article was published on 2026-10-11.
- Where
- The testing began on an M5 Ultra Mac Studio and also involves other Mac and Windows systems.
- Why
- The writer wants to explore local AI while avoiding giving personal data to cloud services.
This story does not have two clearly opposing sides.
What do I do with this thing?
The writer says concerns about giving personal data to cloud services had held them back from using AI regularly.
The writer installed Hermes Agent on an M5 Ultra Mac Studio and selected a 125-billion-parameter Qwen model.
The writer set up a daily morning briefing and made Hermes controllable through a Telegram bot.
- Software
- Hermes Agent, an open-source self-hosted AI agent
- Computer
- M5 Ultra Mac Studio with 256GB of unified memory
- Model
- Qwen 3.8 Flash Next, with 125 billion parameters
- Model size
- Around 105GB
- Messaging service
- Telegram











