lancew/Qwen3-8B-GPTQ-INT8

Benchmark

Capability Benchmark	Qwen3-8B-FP8	Qwen3-8B-INT8
aime25@mean_acc	0.3778	0.3555
gpqa_diamond@mean_acc	0.5572	0.5589
ifeval@mean_prompt_level_strict	0.8182	0.8152
live_code_bench@mean_acc	0.557	0.5703
mmlu_pro@mean_acc	0.7058	0.7118

• 查找"stop_reason": "max_tokens"的sample，发现均为重复回答至max_token，未发现有正常回答至最长的情况
• repeats : 3
• enable_thinking: True
• evalscope version: 1.4.2

Reference

• Best Practices