# The model sizes

**URL:** </static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fthe-model-sizes%2F191258%3E%5C
**Category:** Site Feedback\
**Created:** [October 31, 2026, 7:54am UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fthe-model-sizes%2F191258 "2026-10-31T07:54:17Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![klimtherider](https://avatars.discourse-cdn.com/v4/letter/k/6bbea6/32.png) [@klimtherider](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Fu%2Fklimtherider)%5C
**Post date:** [October 31, 2026, 7:54am UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fthe-model-sizes%2F191258%2F1 "2026-10-31T07:54:17Z")

</div>

Man! We are live in 2026! Where the model size? I understand that you download some 2-10 Gigs, but 200-300 Gigs? And you dont know what size is it? Are you sure is it OK?

---

<div class="post-metadata">

**Author:** ![thesocialsensei](https://avatars.discourse-cdn.com/v4/letter/t/2acd7d/32.png) [@thesocialsensei](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Fu%2Fthesocialsensei)%5C
**Post date:** [November 1, 2026, 2:41am UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fthe-model-sizes%2F191258%2F2 "2026-11-01T02:41:59Z")

</div>

It helps to know where the number comes from: file size is roughly parameters × bytes per weight. In the original 16-bit (bf16) format that’s 2 bytes per parameter, so a 7B model is about 14 GB, a 70B about 140 GB, and the big open models with 100B+ parameters (including mixture-of-experts models) land at 200–500 GB. Nothing is hidden, though: the model card or config.json gives the parameter count, and the “Files and versions” tab lists every file’s size before you download.

For running a model at home you rarely want the original weights anyway. Look for quantized versions (repos usually named …-GGUF). Q4\_K\_M is about 0.6 bytes per parameter, so a 7B is around 4.5 GB and a 70B around 42 GB. A rough rule: you need that much RAM or VRAM plus a few GB for the context.

So yes, 200–300 GB is normal for the largest open models. Those are meant for multi-GPU servers; the quantized repos are the ones aimed at home hardware.