How much memory do you need to run an AI model on your own computer?
By Muhammad Asif · · 2 min read
A practical guide to matching model size to your computer's memory, using real download sizes from the Ollama library.
The basic rule
A model must fit in memory while it runs. The file you download is a good first estimate of how much memory the model needs, and you need extra room for the operating system, your other programs and the conversation itself. A model that is the same size as your total memory will not run well.
Bigger models usually give better answers but need more memory and run slower. Smaller models are fast and fit on ordinary laptops, but they are less capable.
Real download sizes
These sizes come from the Ollama model library, read on 2026-11-17. They are for the default downloads, which are compressed (quantized) versions.
| Model | Size | Notes |
|---|---|---|
| Qwen2.5 0.5B | 398 MB | Tiny, for experiments |
| Gemma 2 2B | 1.6 GB | Small and fast |
| Qwen2.5 3B | 1.9 GB | Small |
| Mistral 7B | 4.4 GB | Popular first choice |
| Qwen2.5 7B | 4.7 GB | Multilingual |
| Llama 3.1 8B | 4.9 GB | General chat |
| Gemma 2 9B | 5.4 GB | Compact |
| Qwen2.5 14B | 9.0 GB | Needs a stronger machine |
| Gemma 2 27B | 16 GB | Workstation class |
| Qwen2.5 32B | 20 GB | Workstation class |
| Llama 3.1 70B | 43 GB | Server or high-end workstation |
A simple starting guide
- 8 GB of memory: stay with models of about 3 billion parameters or less.
- 16 GB: 7B to 8B models are the sweet spot.
- 32 GB: models around 14B to 27B become possible.
- 64 GB or more: the 32B and 70B sizes are possible if enough memory is free.
These are rules of thumb, not guarantees. Speed depends on your processor and graphics card as well.
Longer conversations use more memory
Ollama's default context length is 4096 tokens. If you raise it so the model can read longer documents, it needs more memory. If your computer slows down, lower the context or choose a smaller model.
How to check
Run ollama ps while a model is loaded. It shows whether the model runs on the GPU or the CPU. A model running on the CPU is much slower.
Next steps
Try a small model first, then move up one size at a time. Our hardware guide and model pages list sizes and licenses.

