[v3ss0n]

That is true too. But I found Qwen3 14B with 8bit quant fair better than 32B with 4b quant . Both kvcache at 8bit. ( i enabled thinking , i will try with /nothink)