yeths

/

The-Quettamind

Like 29

and then lying in the first line about 28 mb thing

Huggingface uses Xet storage, meaning despite the fact that, yes, the files appear to equal nearly 10TB, it does in fact only use 28MB on their servers. The model itself consists of pure zeroes, so Huggingface deduplicates it. See https://huggingface.co/docs/hub/xet/deduplication to understand how this works.
Edit: It's also important noting that this is a fun experiment I made, and explicitly not advertised as a frontier model (see the benchmark table lmao)

Anuragog311... its less than 5mb... this is cause hf estimates size based on the paramaters count. Eg, 1b * 2bytes / p (dtype FP16) it will assume 2gb of weights. its not actually counting the disk size. They used the smart option of doing simple math instead of checking sometimes 400+ files, which can take Linux a bit of time. This math falls apart with Xet storage. Typically, models are billions of unique numbers. when they are all 0s xet can compress it heavily. This model is 5mb of metadata basically...

Also CoderX just so you know they patched the thing where you can declare your own paremeter counts or lie about it, they actually check, which is why i have to have so many terabytes of files on this repo. They are all just deduplicated though

I know that. They fixed the metadata bug in gguf files which people kept spamming. But I was tryna argue for u that its not a ultra huge files. I was trying to prove it.

Thanks