About VRAM Calculator
VRAM Calculator tells you how much GPU memory a local LLM needs before you download it or buy a card. It is built and maintained by a small team. We made it because a parameter count is a poor guide on its own: the same model needs several times more memory as a BF16 file than as a 4-bit GGUF, and a long context adds gigabytes on top.
What the calculator does
- Adds up the three things that use memory: the weight file you actually download, the KV cache for your context length, and the runtime's own buffers.
- Compares that total with the memory a GPU, Mac or DGX Spark can really use once the OS and display take their share.
- Covers 86 models and 32 GPUs and unified-memory machines, plus 78 model-on-GPU pages under Can I run it?
- Runs entirely in your browser. Settings live in the page URL, so a link reproduces your result.
What it does not do
- It does not scan your PC or probe your GPU. You pick the hardware.
- It does not download weights, run models or proxy requests to a hosted model.
- It does not predict tokens per second or rank cards by speed or price.
- It estimates inference only. Training and fine-tuning memory are out of scope.
- It does not add two GPUs into one bigger pool, and it does not count a Mac's unified memory twice as both VRAM and RAM.
Where the numbers come from
Model shapes (layers, attention heads, context window) come from each model's official config.json on Hugging Face. Weight sizes are the byte counts of the published files, such as a GGUF Q4_K_M file or the official safetensors shards, not a guess from parameter count. GPU capacities come from the vendor's spec page. Each model and GPU page links its sources.
The totals are estimates. So far we have measured 3 runs on real hardware and compare them with the estimate on the accuracy page. The formulas are written out on the methodology page.
Found a wrong number?
Email us. Tell us the model, the exact file or quant, your GPU, the context length and the runtime with its version. If you have a measured memory reading, include it. The contact page lists what helps for each kind of report.