I'm using official @8bit quants from Qwen, they maintain more capability