In this article 6 sections
- 1. Check where the model is loaded with ollama ps
- 2. Compare the model size with available GPU memory
- 3. Check the default context length of 4096 tokens
- 4. Review the check order all at once
- 5. Limitations of this article
- References
When a local AI model running in Ollama responds slowly, it is best to check its execution status before changing models. Based on explanations in the official Ollama FAQ, this article organizes the checks into whether the model is loaded onto the GPU, how long the context length is set to be, and whether other programs are using memory. It is not a measurement of speed on an actual PC and does not guarantee any response-time figures.
1. Check where the model is loaded with ollama ps
After running the model, enter the ollama ps command in the terminal. You can check where the model is loaded in the table’s PROCESSOR column. 100% GPU means that the entire model is loaded into GPU memory, while 100% CPU means that it is running in system memory. If the values are mixed, such as 48%/52% CPU/GPU, some processing is handled by the CPU, which may make responses slower.
The official FAQ’s example shows the llama3:70b model with SIZE 42 GB and PROCESSOR 100% GPU. This is an example for documentation and is not the result from any particular PC. On your own PC, you must check the SIZE and PROCESSOR values directly using the same command.
2. Compare the model size with available GPU memory
The SIZE shown by ollama ps is the size occupied by the running model. If this value is larger than the available GPU memory, some of the model may be moved to the CPU. In Windows, you can view GPU memory usage in the Performance tab of Task Manager. If games, browsers, or video-editing programs are using a lot of GPU memory, close them and try running the model again.
Reducing the model size is another option. The llama3.2 page in the Ollama library lists 1B and 3B sizes, with 3B as the default tag. Choosing a smaller size may change the loading status, but it may also change answer quality, so it is a good idea to compare the two sizes using the same question.
3. Check the default context length of 4096 tokens
The official FAQ states that Ollama uses a default context window of 4096 tokens. If you paste in a long document or continue a conversation for a long time, the beginning may be cut off when this range is exceeded. To change the value, set the OLLAMA_CONTEXT_LENGTH environment variable or specify the /set parameter num_ctx value inside ollama run. If you use the API, set num_ctx inside options.
Increasing the context value allows more content to be read at once, but it may also increase memory usage. Unless your purpose is summarizing long documents, it is safer to start with the default value and increase it only when necessary.
4. Review the check order all at once
1) Check the PROCESSOR value in ollama ps. 2) Compare SIZE with the available GPU memory. 3) Check whether num_ctx or OLLAMA_CONTEXT_LENGTH is larger than necessary. 4) Close demanding programs and run the model again. 5) If it is still slow, compare the same question using a smaller model size.
If you change several values at once, you will not know what was effective. Change one item at a time and note the changed value and response time to make it easier to separate the causes.
5. Limitations of this article
This article summarizes explanations from Ollama’s documentation. How fast a particular model runs on a particular GPU varies depending on the PC specifications, drivers, and model type, so this article does not predict it. For items to check before installation, see Specifications to check before installing Ollama on Windows ↗, and for the difference between GPU memory capacity and bandwidth, see AI GPU memory ↗.
References
- Official Ollama FAQ (default context length, environment variables, and interpreting ollama ps output): Ollama FAQ ↗ (checked 2026-11-11 KST)
- Ollama library llama3.2 page (1B·3B sizes and default tag): llama3.2 ↗ (checked 2026-11-11 KST)
Sources and original article
Read the Korean original on Naver ↗
This article reflects public information available when it was written. Specifications, prices and schedules can change; check the linked official documentation.
This English edition was written and translated with AI assistance, with sources provided. For errors or additional information, please request a correction. Editorial policy