Ollama Making Your PC Slow? What to Check First

Ollama starts and the model answers, but your computer suddenly feels unusable. Start by identifying the symptom: is the whole system slow, every response slow, or only the first request slow? Each case calls for a different check.

A September 4 user report describes poor PC performance while running a local model. That experience alone does not identify whether RAM, GPU placement, context length, or background applications caused the slowdown.

1. The whole computer slows down

Open Task Manager or Activity Monitor and observe memory use during a request. Check which other applications are consuming significant memory. Compare the same short prompt after closing applications you do not need.

Write down the result. “It feels better” is a useful clue; recording the duration of the same test makes comparisons easier. Change one factor at a time so you can tell which adjustment helped.

2. Every response is slow

Run this read-only command while the model is loaded:

ollama ps

The Ollama FAQ explains that PROCESSOR shows whether the model is loaded on the GPU, CPU, or split between them. CPU execution is possible. If you expected an available GPU to be used, check the hardware documentation. Adding RAM alone does not enable GPU acceleration.

3. Long conversations become especially slow

A larger context window requires additional memory. Use the Local AI Hardware Checker to compare, for example, 8k and 32k context while keeping model size and quantization unchanged. The calculator provides a planning estimate, not measured throughput.

Ollama's context-length documentation explains the relationship. Do not automatically choose the smallest window: a task involving long documents may lose relevant information if the usable context is too short. Choose a size that still supports the actual task.

4. Only the first request is slow

Compare two short requests to the same model. If the second is much quicker, initial loading may explain the difference. Use ollama ps to check whether the model remains loaded between requests. This is a diagnostic clue, not proof of a storage bottleneck.

Before buying hardware

Start with the symptom selector in the Hardware Checker. Then compare a smaller model, an appropriate quantization level, or a shorter context. Keep the model name, quantization, context, prompt, and response duration in your notes. A repeatable experiment is more useful than an upgrade based on one frustrating run.

Source check — September 7, 2026: The linked Ollama documentation explains functionality; the X post supplies the problem context. Actual speed depends on your full setup.