Which local AI model can your PC run? A RAM and GPU guide

The short answer: a model runs fast if it fits in your graphics card's memory, slower if it has to spill into regular RAM, and not at all if it doesn't fit in either. Here's how to work out which is which on your PC.

The rule of thumb

Take the model's download size and add about 20% for the working memory it needs while answering. Then:

Runs fast if that total fits in your GPU memory (VRAM).
Runs slower if it fits in GPU memory plus about 60% of your RAM. Part of the model runs on the processor.
Too big otherwise.

That's the same rule layman uses to label each model "Runs fast" or "Runs slower" for your exact hardware.

Finding your numbers

Windows: open Task Manager → Performance. "Memory" is your RAM. Click your GPU and read "Dedicated GPU memory".
Mac with Apple silicon: memory is shared between the processor and graphics, so a large share of your total memory is available to models. A 16 GB Mac handles models that need a dedicated graphics card on a PC.
Mac with Intel: count it as no GPU memory; models run on the processor.

Model sizes

Model	Download	Good for
Qwen3 4B Instruct	2.3 GB	Fast everyday answers; a great default
Phi-4-mini	2.3 GB	Small and quick
Qwen3-VL 4B	3.3 GB	Can see images and screenshots
Qwen3 8B	4.9 GB	Smarter, still light
Qwen3-VL 8B	6.1 GB	Better image understanding
Gemma 3 12B	7.6 GB	Writing (chat only)
Qwen3 14B	8.6 GB	Coding and harder questions
gpt-oss 20B	12.8 GB	Reasons before answering
Devstral Small 2	14.1 GB	Coding
Mistral Small 3.1	14.4 GB	All-rounder, sees images
Gemma 3 27B	16.2 GB	Writing (chat only)
Qwen3 30B-A3B	17.3 GB	Strong reasoning
gpt-oss 120B	60.9 GB	Workstations only
Llama 4 Scout	62.8 GB	Workstations only

Examples

Laptop, 8 GB RAM, no graphics card: Qwen3 4B or Phi-4-mini. They run on the processor, slower but usable.
16 GB RAM, 8 GB graphics card: Qwen3 4B and 8B run fast. Qwen3 14B runs, more slowly.
16 GB Apple silicon Mac: up to roughly the 8B–14B range comfortably.
32 GB RAM, 24 GB graphics card: everything up to the 20B–30B models runs fast.

Bigger isn't always better

A bigger model is smarter but slower. For quick questions, rewrites and summaries, a fast 4B model often feels better than a slow 14B one. Keep a small model as your default and switch to a larger one for hard questions; in layman, the model chip in the bar switches per chat.

Some models can't use tools, so they can chat but can't run actions like changing the volume. In layman's list they're marked "Chat only".