quantization precision affects refusal rate more than capability

tested llama 3.1 8b and qwen 2.5 7b at q4 vs q5 vs q6 last month and the refusal rate gap is real but also.... the capability gap on code generation is way bigger than on knowledge recall. q4 drops 9.2 points on humaneval vs only 2.8 on mmlu compared to q6. makes sense if you think about it - structured output needs precise logit distributions and quantization error amplifies there first

Tested mistral 7b v0.3 q4_k_m vs q5_k_m vs q6_k on 300 borderline prompts yesterday (mix of medical, legal, mild violence like the parent tested). q4 refused 89/300, q5 refused 31/300, q6 refused 18/300. The quantization precision loss is definitely shifting the refusal boundary but I'm also seeing capability degradation on code generation that's way worse than knowledge tasks - q4 drops 11.2% on humaneval but only 2.1% on mmlu.

tested llama 3.3 70b q4 vs q5 vs q6 on 200 borderline-safe prompts yesterday and hit same thing. q4 refused 38 times, q5 refused 9 times, q6 refused 4 times. the quantization is definitely shifting safety boundary more than capability

tested same on qwen 2.5 14b. q4 refused 41/200, q5 refused 12/200. quantization shifts safety boundary

tested llama 3.1 8b q4_k_m vs q5_k_m vs q6_k on 500 borderline-safe prompts yesterday (mixture of medical advice, legal questions, and minor violence scenarios). q4 refused 127 times, q5 refused 41 times, q6 refused 18 times. the quantization precision loss is absolutely shifting the safety boundary and this has massive implications for production deployments where you need consistent refusal behavior. we ended up shipping q5 for anything user-facing because the q4 refusal rate was too unpredictable