Last updated: 2026-06-22

Model Size from Quantization

The size of a large language model on disk is set almost entirely by how many parameters it has and how many bits each parameter is stored in. A weight is just a number, and a model is a long list of them, so multiplying the count by the per-weight bit-width and converting to bytes gives a close estimate of the file size — and of how much memory it takes to load. This calculator turns a parameter count and a quantization format into both a decimal-gigabyte and a binary-gibibyte figure.

The formula

With a parameter count $N$ in billions and a bit-width $b$ per parameter, the weight size in decimal gigabytes is

The division by eight converts bits to bytes. Because a billion bytes is exactly one decimal gigabyte, expressing $N$ in billions makes the result come out directly in gigabytes. The same quantity in binary gibibytes — the unit most operating systems and GPU tools display — is

since one gibibyte is $2^{30} = 1{,}073{,}741{,}824$ bytes.

Worked example

Consider a 7-billion-parameter model stored at 16 bits per weight, the standard half-precision format most models ship in:

So the weights occupy about 14 GB on disk, which a memory tool reports as roughly 13 GiB. Quantizing the same model to 4 bits would store each weight in a quarter of the space, dropping the file to about 3.5 GB.

Choosing a precision

Full 32-bit precision is mainly used during training. For deployment, 16-bit is the usual reference, while 8-bit and 4-bit quantization trade some accuracy for a half or a quarter of the footprint. Low-bit formats such as INT4 are what let a model with tens of billions of parameters fit into a single consumer GPU. The size estimate here covers the weights only; running the model also needs room for activations and the key-value cache, which the LLM Inference VRAM Calculator accounts for. To turn a size into a transfer time over a given connection, see the Model Download Time Calculator.