GGUF Memory Estimator

GGUF Memory Estimator estimates the VRAM and RAM a GGUF model needs, given only a Hugging Face repository path, by fetching the model's metadata through HTTP Range requests rather than downloading the model file itself. It runs in a browser, served over local HTTP because the CORS headers it needs are blocked under the file protocol, or from the command line with options for context length, batch size, separate key and value cache quantization types, and available VRAM or RAM to add a fit check to its output.

Its calculation covers model weight size computed from exact per-quantization-type byte sizes in the GGUF metadata, key-value cache size from layer count, head count, head size and context length, activation memory including the mixture-of-experts case, and multimodal projector weights and their per-image activation cost. Performance flags can add a decode, prefill and time-to-first-token estimate for a named GPU or CPU, and mixture-of-experts flags can move expert weights between VRAM and RAM.

It is written in JavaScript and runnable through a local Python HTTP server or a Node command-line script.