LLM Sizing Calculator: will it run on your machine?

What the numbers mean

• Weights is parameters × bytes per parameter, or the real file size if you enter it or pick a GGUF file. Some models ship big embedding tables or image and audio encoders, so the download can be larger than the estimate.
• KV cache is 2 × layers × KV heads × head dimension × tokens × bytes per value. Models with sliding-window or shared-cache layers, like Gemma, need less than this.
• Usable memory is the full amount for VRAM and RAM, and 70% of unified memory on a Mac, the GPU's default share.
• Speed is 0.6 × bandwidth ÷ GB read per token. For Mixture-of-Experts models only the active parameters are read, so they run at small-model speed while still needing memory for all of them.

The Hugging Face lookup reads the parameter count, config.json, and GGUF file sizes straight from Hugging Face's public API. Gated models such as Meta's Llama don't publish config.json without a login, so the lookup reads it from an ungated copy of the same model when one exists. If it can't find one, the KV cache is a rough estimate from the parameter count, which can be off by 2× or more, until you enter layers and heads from the model card. For a Mixture-of-Experts model, the active parameter count is read from names like "30B-A3B"; otherwise, enter it yourself.

These are estimates, usually within 20–30% of reality. The calculator covers generation speed only; reading a long prompt first is limited by compute, where a GPU is often 10× faster.

For the reasoning behind each formula, read Will It Run? How to Tell If an LLM Fits on Your Computer or take the free LLM Sizing 101 course.