LGAI-EXAONE

/

EXAONE-4.0-32B-AWQ

like 12

Follow

LG AI Research 1.66k

That's because it is AWQ Quants model, and HF calculate params as like that (almost 1/4 of original fp16 model).
Other AWQ models also shows like that. ( for example, https://huggingface.co/Qwen/Qwen3-32B-AWQ )

Thanks for the note, I'll close this out, but wondering if that value is accurate? i.e. does the AWQ quantizing actually reduce the parameter count? Or is just weird maths used by HF on the card pages (i.e a known bug) ?