Last updated: 2026-06-22

Model Download Time

Pulling a large language model onto a machine can mean moving tens or hundreds of gigabytes, and the wait depends on two things: how big the weight file is and how fast the connection runs. The file size follows from the parameter count and the quantization format, while the transfer time is that size divided by the network bandwidth. This calculator chains the two together, starting from a parameter count and a precision rather than asking for a file size directly.

From parameters to bytes to time

A model's weights occupy roughly its parameter count times the per-weight bit-width, in bytes. With $N$ parameters in billions and $b$ bits each, the size in decimal gigabytes is

To turn that into a transfer time, the size has to be converted to bits — because bandwidth is quoted in megabits per second — and then divided by the link rate. One gigabyte is $8 \times 10^9$ bits, and one megabit per second is $10^6$ bits per second, so over a bandwidth $w$ in Mbps the transfer takes $S \cdot 8 \times 10^9 / (w \times 10^6) = S \cdot 8000 / w$ seconds. Expressed in minutes and hours,

Worked example

Take a 70-billion-parameter model at 16 bits per weight, downloaded over a 1,000 Mbps connection:

That is about 0.31 hours of transfer for the 140 GB of weights, assuming the link stays saturated. The same model quantized to 4 bits would be 35 GB and download in roughly a quarter of the time.

Megabits versus megabytes

The most common mistake is forgetting that connection speeds are in megabits while files are in megabytes. A byte is eight bits, so a "1 Gbps" link moves only about 125 megabytes per second. The calculation above handles this conversion, which is why the time is eight times longer than dividing gigabytes by gigabits would suggest. Real downloads are slower still, because server limits, shared bandwidth, and protocol overhead keep the effective rate below the advertised maximum. The underlying file size, and how it changes with precision, is covered in the Model Size from Quantization Calculator, and the memory needed to actually run the downloaded weights in the LLM Inference VRAM Calculator.