Local AI sizing tool
LLM RAM and VRAM Calculator by Model, Quantization and Context
Estimate the weight-memory floor for an LLM, then add explicit runtime and context reserves. This calculator deliberately does not use a generic KV-cache formula because cache memory depends on the model architecture, runtime, data type, batch size, and context configuration.
Calculation method
Weight memory = parameters × bits per parameter ÷ 8. Runtime reserve is then added as a percentage of weight memory, followed by the explicit context reserve you enter.
The result is intentionally conservative and transparent. Actual runtime memory can differ because of model architecture, KV cache, batch size, framework overhead, temporary buffers, and implementation.
How much RAM or VRAM does a local LLM need?
The minimum starts with model size and quantization, but real runtime memory is higher. Context length, framework overhead, temporary buffers, batch size, and the exact model architecture all affect the final requirement.
Is 16GB of VRAM enough for a local LLM?
It can be enough for many smaller models and some larger quantized models, but it is not a universal cutoff. Leave headroom rather than assuming a model is safe simply because its weight file is smaller than 16GB.
Why this calculator uses a manual context reserve
Exact attention-cache memory cannot be determined reliably from parameter count alone. Entering an explicit reserve keeps the result honest while still making the tool useful for capacity planning.