Hermes Agent runs local Qwen AI models on Mac hardware
Recent testing shows open-source tools like Hermes Agent can locally process complex tasks using massive LLMs, offering privacy for sensitive data without relying on cloud subscriptions.
As local hardware capabilities expand, developers and power users are increasingly running large language models directly on personal devices. Using the free, open-source Hermes Agent desktop application on macOS, Windows, and Linux, practical experiments demonstrate how locally hosted systems offer an appealing, privacy-focused alternative to paid cloud subscriptions managed by tech giants like OpenAI, Google, Microsoft, or Anthropic.
On an M5 Ultra Mac Studio equipped with 256GB of unified memory—a $12,000 setup—the Hermes setup successfully loaded Qwen 3.8 Flash Next, a 125-billion parameter model spanning approximately 105GB. Smaller configurations are also being tested across hardware including an M6 Mac Mini, an M5 MacBook Air, an Asus TUF Gaming A14 powered by an AMD Strix Halo processor, and upcoming RTX Spark Windows machines featuring up to 128GB of RAM. In extreme enterprise demonstrations, Apple even showcased a cluster of four Mac Studios linked together, representing roughly $50,000 in local compute power.
Deploying these local models enables automated workflows without sending sensitive data to external servers. By granting access via a Telegram bot and a Steam web API key, Hermes organized a library of over 400 games by genre. It also processed sensitive financial records, created confidential laptop spec spreadsheets, and attempted to generate Python scripts to automate benchmark tests that normally require running three manual passes to calculate averages. Additionally, it executed a scheduled 7:30am cron job to compile daily briefings from email, calendar, and weather data.
For AI practitioners, these experiments show that offloading tasks to local agents is technically viable and safeguards sensitive information under embargo or strict privacy rules. However, local execution still faces practical hurdles, including system sleep issues interrupting scheduled tasks and inconsistent outputs, proving that local hardware can handle heavy computational lifting even if autonomous reliability remains a work in progress.
This is our own summary of reporting by The Verge AI
