Start a prompt, open a second Terminal window, and run ollama ps while the model is loaded. Read the PROCESSOR column. Ollama reports whether the model is loaded on the GPU, on the CPU, or split between them.

Read the result before changing hardware

100% GPU: The model is loaded entirely on the GPU. If it still feels slow, compare the same prompt with a smaller model before drawing conclusions about the Mac.

CPU/GPU split: Keep the model fixed and retry with a shorter context, then compare against a smaller model. Change one variable at a time.

100% CPU: Confirm you are checking the intended Ollama instance and review the platform and GPU support notes.

Separate speed from memory pressure

Open Activity Monitor → Memory while the request runs. If memory pressure rises when you feed the model a long file, compare with a shorter input using the same model. Repeat with the same prompt and settings so you can tell which change helped.

Ollama’s FAQ explains the processor column and context settings.

Run a comparison you can repeat

Start with a fresh chat and a short prompt that produces a predictable answer, such as summarizing the same paragraph in three sentences. Repeat once after the model is loaded so you can distinguish startup delay from slow generation. Record the model tag, context setting, processor placement and how long you wait before text starts appearing. Then repeat with a smaller model or shorter input. Keep the other settings the same.

What a real result might tell you

Imagine a model answers short prompts comfortably but slows down when you paste a long log. If Activity Monitor also shows rising memory pressure, try sending only the relevant excerpt and rerun the comparison. If a short prompt is still slow and the processor column reports a CPU/GPU split, try a smaller model or a lower-memory variant and check placement again.

A model showing 100% GPU can still be slow. If the first response is the only slow one, compare the next prompt before changing hardware. If all responses are slow, test a smaller model using the same prompt and judge whether its answer is good enough for your task. For a remote setup, repeat the check on the Mac actually running Ollama. The client laptop’s Activity Monitor describes the client, while the model is running elsewhere.