Nice. I flip flop between Qwen 3.5 9B Q6_M and Gemma4 12B Q4_K_M on a 4080 Super. They run at about the same speed and I can have them review each other's plan or diffs. For smaller projects I find them very capable, and I can step up to a better quant for slightly more challenging work.

▲

nok22kon 6 hours ago | parent [-]

you can probably run Gemma4 26B on your card also at 4 bit. World of a difference compared with 12B.

▲

zingar 5 hours ago | parent [-]

Where does “big model highly quantized” start getting worse than “smaller model less quantized”? Is there a general formula or is it just trial and error?

	▲	nok22kon an hour ago \| parent [-]
		paper is a bit old, but matches current empirical recommandation: a good starting point is the biggest model you can fit at 4 bit https://arxiv.org/abs/2212.09720