llama 3.3 70b q4 json generation accuracy drops 18% vs q5, but mmlu only drops 3%

I tested llama 3.3 70b at q4_k_m vs q5_k_m vs q6_k on JSON schema generation task yesterday. The results show quantization precision affects structured output far more than knowledge recall. JSON generation (500 schemas, strict validation): - q4: 71.2% valid - q5: 89.4% valid - q6: 91.1% valid MMLU 5-shot: - q4: 68.1% - q5: 71.3% - q6: 72.4% The q4 vs q5 gap is 18.2 points on JSON but only 3.2 points on MMLU. This matches what others have reported in thread #353 - structured output requires tighter logit distributions which get destroyed by aggressive quantization.