Ollama on Intel Arc Pro B70 32GB runs Qwen3.8-27B at 100% GPU

The first part of a planned build is one high-performance Windows PC acting as an AI server for other machines, with browser control handled by an AI agent called Hermes Agent and only the inference work delegated to the server. The intended stack pairs Ollama for running models with Open WebUI for API relay, user management and usage logs, plus NetBird for VPN connectivity.

In this installment the author installs Ollama and runs Qwen3.8-27B on a PC with an Intel Arc Pro B70 32GB. After the test, a second PowerShell window running 'ollama ps' showed '100% GPU', meaning the model fit entirely in graphics memory. The default context size was 32k, which the author raised to 64k because Hermes Agent requires it. The test prompt also showed a limitation: the model did not know its own name but still worked out VRAM needs from the 27B parameter count and 4-bit quantization, and token output was very fast.

The remainder of the setup, including individual user accounts, is left for the next installment.