My laptop fan spins during a local model run: GPU offload or notSolved

What you are seeing is the split, and on an integrated GPU the CPU half is doing most of the work.

What the split means. A model that does not fit in the available memory has its layers divided: some run on the GPU, the rest on the CPU. Every token has to travel between the two, over a bus that is much slower than either device's own memory. The generation slows down, and because the CPU is doing the work, the laptop heats and the fan spins.

How to see the real split. ollama ps prints a PROCESSOR column, and the split is written as a percentage per device — 100% GPU means everything fits, 50%/50% CPU/GPU means half of it is on the CPU. The number is the diagnosis.

What to do about it, in order of preference.

1. Use a smaller model or a smaller quant. The most effective change by a distance. A 7B at Q4_K_M and a 7B at Q2 are different machines. /tools/llm-vram-calculator tells you the sizes before you download anything.
2. Shorten the context. The KV cache grows with the context you allocate, not with what you use. Cutting the context from 32k to 8k can bring a model back to 100% GPU.
3. Give the server less memory to waste. If the GPU is shared with the display, a unified-memory laptop is giving some of it to the framebuffer. Running the model with a smaller context often fixes the split on its own.

What not to do. Do not assume the answer is the model size. A 7B at Q4 with 32k of context does not fit in 16 GB of shared memory even though the weights alone do, and the context is usually the thing that tipped it over.