I ran local LLMs on Intel's cheapest iGPU, and the results were surprisingly decent

A hands-on experiment running local LLMs on an Intel N100 compute module (LattePanda Mu) using llama.cpp inside a Proxmox LXC container with iGPU passthrough. The setup involved compiling llama.cpp with Vulkan support, troubleshooting RAM constraints during compilation, and testing models including Gemma 3 4B, Qwen3 4B, and DeepSeek R1-Distill-Qwen-7B. The N100 managed ~2.9 tokens/second on the 7B model, making it viable as a lightweight secondary LLM server or embedding model host, though not a replacement for a dedicated GPU setup.