6-bit weights beat extra passes

🔥 What happened Researchers tested Qwen3-0.6B and 1.7B with simulated weight quantization (8, 6, 4, 3 bits) and option-order averaging. 6-bit weights match fp16 accuracy — even score slightly higher. At 4 bits, answer-slot skew jumps sharply. 💡 Why it matters Averaging four cyclic orders recovers +12.6 points at 4 bits for the 0.6B model. But cost-matched, extra orders only beat extra bits at 6+ bits. The decision table: use 6-bit weights and spend the rest on up to four orders, one pass each. ⚡ Our take Saving bits at 4-bit precision buys you order-skew debt you'll pay for later. 6 bits is the sweet spot — everything else is tinkering.