GPU Picker
This is a framework that teaches the decision, not a shopping list — it will never name a product or quote a price. Cards age; VRAM tiers do not. Answer for the models you actually want to run, how hard you will quantize, and your budget posture, and it resolves to a card class — then the sizer turns your exact model into an exact requirement.
Question 1
What size of model do you want to run locally?
Sizes are parameter counts, not products: small ≈ up to 8B parameters, mid ≈ 13–34B, large ≈ 70B and up.
How to read it: VRAM is the only spec that decides whether a model runs at all — everything else decides how fast. That is why the questions are about bytes: model size sets the parameters, quantization sets the bytes per parameter, and their product is the card class you need. Notice how often the answer to a bigger ambition is not a bigger card but a harder quantization — that arithmetic is the cheapest upgrade in local AI.