オープンウェイトモデル

Qwen 3 32B

Large dense reasoning model; Q4 is tight on 24GB once KV cache and runtime overhead are included.

既定コンテキストでの量子化メモリ

量子化	重み	KV cache	合計VRAM
FP16	66.00GB	2.15GB	74.75GB
Q8_0	36.00GB	2.15GB	41.75GB
Q6_K	27.95GB	2.15GB	32.89GB
Q5_K_M	23.65GB	2.15GB	28.16GB
Q4_K_M	20.40GB	2.15GB	24.59GB
Q3_K_M	16.90GB	2.15GB	20.74GB
Q2_K	10.75GB	2.15GB	13.97GB