LLM VRAM calculator: can my GPU run it?

How it works

Running a model locally needs memory for three things: the weights, the KV cache that holds the conversation, and the software itself. The calculator works out each one.

• Weights = parameters × bits per weight ÷ 8. Bits per weight for each quantization come from llama.cpp’s own measurements.
• KV cache = 2 × layers × KV heads × head size × context length × bytes per value × chats running at once. Layers, KV heads and head size come from each model’s official configuration file. For gpt-oss, half the layers use a 128-token sliding window, and the calculator counts them that way.
• Runtime: an allowance you can change, because it depends on the software you use.

The speed estimate

When a model writes, each new token reads every active weight from memory once, so memory bandwidth sets a ceiling: bandwidth ÷ size of the active weights. Mixture-of-experts models such as gpt-oss and Qwen3-30B-A3B read only their active experts per token, which is why they can be fast for their size. Real speed is lower than the ceiling. Where a GPU maker does not publish bandwidth, the tool says so rather than estimating it.