Developer & Tech

Model Size Calculator

What this does

Compute on-disk model weight size in decimal GB and binary GiB, plus download time over a 100 Mbps link.

Enter your details

Runs in your browser

Calculator inputs

How to use it

Using the model size calculator

  1. 01

    Enter parameter count

    Straight off the model card.

  2. 02

    Choose the distribution dtype

    Most open models ship fp16/bf16 safetensors; GGUF quants are int4/int8 equivalents.

  3. 03

    Plan the transfer

    The 100 Mbps row sets expectations before you kick off large pulls.

Good to know

GB versus GiB, made concrete

A “140 GB” checkpoint occupies 130.39 GiB on disk. Tools reporting binary units while marketing quotes decimal create recurring confusion; this calculator shows both so a df listing and a spec sheet finally agree.

Sharded checkpoints

Large models ship as many safetensors shards whose sizes sum to the figures here, plus tokenizer and config JSONs worth kilobytes. Training resumes pull optimizer state files several times larger; check what a repo actually contains before mirroring it.

How it's calculated

The math behind this calculator

bytes = params × 10⁹ × bytes/param GB = bytes ÷ 10⁹ GiB = bytes ÷ 2³⁰ download s = bytes × 8 ÷ 10⁸

Weights dominate model file sizes: parameters times bytes per parameter. Decimal GB suits marketing sheets and network math; binary GiB suits disks and RAM, differing by ~7%; both shown so units never silently mix. Download time divides total bits by an advertised 100 Mbps decimal link.

Assumptions & limitations

  • Weights only; tokenizer/config files are negligible.
  • Optimizer states in training checkpoints dwarf these figures.
  • Download assumes sustained 100 Mbps with protocol overhead ignored.

Worked example

A 70-billion-parameter model at fp16 weighs 140 GB decimal (≈130.39 GiB); about 3h 06m 40s to pull over a sustained 100 Mbps connection.

FAQ

Frequently asked questions

Does the download include multiple formats?
No; one copy of the selected dtype. Downloading fp32 and int4 versions doubles and quarters respectively.
Why 100 Mbps fixed?
It is a common fast-line reference; scale linearly; a 1 Gbps line downloads ten times faster than shown.
Are GGUF Q4 files exactly half of Q8?
Roughly; mixed-precision quantization makes exact ratios vary a few percent either way.

Keep exploring

Related calculators