Short answer: 8GB of RAM runs small 3B to 4B models, 16GB is the comfortable floor for 7B to 8B models, 32GB opens up 14B and the efficient 20B to 30B models, and you need 48GB to 64GB of fast memory before 70B class models make sense. If you have an NVIDIA or AMD GPU, the model runs fastest when the whole thing fits in VRAM.
This page gives you the numbers we actually use when picking hardware, based on the download sizes Ollama lists for each model. No guesswork about "it depends" without telling you what it depends on.
The rule of thumb
Ollama's default downloads are 4-bit quantized (usually Q4_K_M). At that level a model needs roughly 0.6GB of memory per billion parameters, plus extra for the context window (the KV cache) and whatever else your computer is doing.
- Model file size: listed on each model page at ollama.com/library.
- Add 1GB to 3GB for context at default settings. Long contexts (32K tokens and up) can add several more GB.
- Leave 4GB to 6GB free for Windows, your browser, and Open WebUI if the model is running from system RAM.
Memory needed by model size
Download sizes below are as listed on ollama.com at the time of writing for the default 4-bit tags. "Comfortable" means the model plus context plus normal desktop use.
| Model (Ollama tag) | Download size | Comfortable RAM (CPU only) | VRAM to fit fully on GPU |
|---|---|---|---|
| qwen3:4b | 2.5GB | 8GB | 4GB to 6GB |
| llama3.1:8b | 4.9GB | 16GB | 8GB |
| qwen3:8b | 5.2GB | 16GB | 8GB |
| qwen3:14b | 9.3GB | 24GB to 32GB | 12GB |
| gpt-oss:20b | 14GB | 32GB | 16GB |
| qwen3:30b (MoE) | 19GB | 32GB | 24GB |
| qwen3:32b | 20GB | 32GB to 48GB | 24GB |
| llama3.3:70b | 43GB | 64GB | 48GB (two 24GB cards) |
| gpt-oss:120b | 65GB | 96GB to 128GB | 80GB class |
RAM vs VRAM: which one matters?
VRAM is the memory on your graphics card. If the full model fits there, you get fast replies, often several times faster than CPU only. If it does not fit, Ollama splits the model between GPU and system RAM. That still works, but speed drops toward CPU levels as more of the model spills over.
System RAM is the fallback and the only option on most laptops and mini PCs without a discrete GPU. Here, memory bandwidth sets your speed. Dual channel DDR5 is fine for 7B to 8B models and slow for 30B dense models. Machines with wide unified memory (Apple silicon, AMD Ryzen AI Max) sit in between and can hold much bigger models.
How to check what your model is actually using
- Run a model:
ollama run qwen3:8b - In a second terminal, run
ollama ps. The PROCESSOR column shows100% GPU,100% CPU, or a split like48%/52% CPU/GPU. - A split means the model does not fit in VRAM. Pick a smaller model or a smaller quantization, or shorten the context.
What to buy if you are short on memory
Upgrade RAM first (cheapest fix)
If your desktop or laptop has socketed RAM, going from 16GB to 32GB is the best value upgrade for local AI. Check your exact model's manual for DDR4 vs DDR5 and SO-DIMM (laptop/mini PC) vs DIMM (desktop) before ordering.
- 32GB DDR5 SO-DIMM kits (laptops and mini PCs) on Amazon
- 32GB DDR5 desktop kits on Amazon
- 32GB DDR4 desktop kits on Amazon
Skip this if your RAM is soldered (many thin laptops, Apple silicon, Ryzen AI Max mini PCs). Then the only path is a different machine.
Add a GPU with 12GB to 16GB of VRAM
For a desktop tower, a card with 12GB or 16GB VRAM moves 8B to 14B models fully onto the GPU. See our best budget GPU for Ollama breakdown.
Or get a box built for it
If you want something small and quiet, read best mini PC for local AI. If you are on 16GB today, our best models for 16GB RAM list shows what runs well right now.
FAQ
How much RAM do I need for Ollama?
16GB is the practical minimum for 7B to 8B models like llama3.1:8b or qwen3:8b. 8GB works only for small 3B to 4B models. 32GB lets you run 14B to 30B class models.
Can Ollama run with 8GB RAM?
Yes, with small models such as qwen3:4b (about 2.5GB download). Close your browser tabs, and expect slower replies than on a GPU.
How much VRAM do I need for a 7B or 8B model?
About 8GB of VRAM fits a 4-bit 8B model with a normal context window. 6GB cards will usually split it between GPU and CPU.
How much memory do I need for a 70B model?
A 4-bit 70B model like llama3.3:70b is about a 43GB download, so plan on 64GB of system or unified memory, or 48GB of combined VRAM across GPUs.
Does Ollama use RAM or VRAM?
Both. It loads as much of the model as fits into GPU VRAM and runs the rest from system RAM. Run ollama ps to see the split.