Qwen3 0.6B quant benchmarks prove Q5 optimal

These community tests provide a rare empirical look at how aggressive quantization degrades sub-1B parameter models, confirming Q5 as the ideal balance.

• –GSM8K math performance collapses from 44.9% at Q8 to just 3.1% at Q2, proving logic is highly sensitive to compression
• –IFEval instruction following degrades from 18.7% to 12.9%, showing formatting constraints are slightly more resilient than pure reasoning
• –MMLU knowledge scores remained completely flat at 22.9% across all quants, suggesting the model's factual recall floor is reached immediately
• –The marginal 10 tok/s speed difference between Q2 and Q8 is noticeable but likely not worth the massive intelligence tax