Methodology and key facts
This page explains where every number on VRAMNeeded comes from, so you can check it yourself.
The fit rule
VRAM needed ≈ weights file size + KV cache for your context + runtime overhead.
- Weights are the real download size of the quantized file on Hugging Face (for split files, every shard is added up). We do not estimate from the parameter count.
- KV cache is calculated from the model's own
config.json: 2 × layers × KV heads × head dimension × bytes per value, per token. It is multiplied by the context length you pick. Long contexts can need as much memory as the weights do. - Overhead for the runtime and activations is shown separately on each page, so you can see how much headroom you are relying on.
Where the data comes from
| Figure | Source |
|---|---|
| Parameters, layers, context length | Official model card and config.json |
| File size per quantization | Hugging Face file listing (exact bytes) |
| KV cache per token | Calculated from config.json, formula shown on each model page |
| GPU and Mac memory | NVIDIA, AMD and Apple spec pages |
| Rental prices | Each provider's public pricing page on the "checked" date |
| Gaming VRAM | Published vendor guidance and benchmarks, cited per page |
If a number cannot be traced to a source, the cell says "not yet sourced". We never fill gaps with guesses.
Dates and updates
Every page shows a Last checked date. Prices drift quickly, so always confirm on the provider's page before you rent or buy. When we correct a number, we update that date.
Money and independence
Some links to GPU clouds and stores are affiliate or referral links, labelled "Affiliate partner" next to the link. Commissions never change a fit verdict or a number. Full details are on the Disclosure page.
Use the data
Download the data: models (CSV) · models (JSON) · GPU prices (JSON). Every row carries its source links and checked date.
Found an error? Email vramneeded.partners [at] gmail.com. See also: About · FAQ · VRAM table.