Python 中文交流群

难绷，我昨天测了一下，bf16的版本效果这样 Intelligence Benchmark Comparison Mode Sampled Qwen3.6-35B-A3B-bf16 Qwen3.6-35B-A3B-nvfp4 ----------------------------------------------------------------------------------- MMLU Sample 500/14042 88.6% 88.4% HELLASWAG Sample 200/10042 93.0% 93.5% TRUTHFULQA Full 817 91.2% 89.1% HUMANEVAL Full 164 95.7% 93.9% MBPP Sample 200/500 92.0% 88.5% LIVECODEBENCH Sample 100/1055 46.0% 34.0% --- Detail --- Model: Qwen3.6-35B-A3B-bf16 Benchmark Accuracy Correct Total Time(s) Think -------------------------------------------------------------- MMLU 88.6% 443 500 8361.4 Yes HELLASWAG 93.0% 186 200 2609.3 Yes TRUTHFULQA 91.2% 745 817 10702 Yes HUMANEVAL 95.7% 157 164 5499 Yes MBPP 92.0% 184 200 6927.8 Yes LIVECODEBENCH 46.0% 46 100 19808.4 Yes Model: Qwen3.6-35B-A3B-nvfp4 Benchmark Accuracy Correct Total Time(s) Think -------------------------------------------------------------- MMLU 88.4% 442 500 2460.5 Yes HELLASWAG 93.5% 187 200 678.8 Yes TRUTHFULQA 89.1% 728 817 3223.2 Yes HUMANEVAL 93.9% 154 164 1655.8 Yes MBPP 88.5% 177 200 2913.9 Yes LIVECODEBENCH 34.0% 34 100 5041.3 Yes