deepseek v3 vs qwen 2.5 coder 32b - tested on 8 benchmarks, results don't match the hype

Everyone's hyping deepseek v3 as the new code generation king, so I spent the weekend running it against qwen 2.5 coder 32b on 8 different benchmarks (HumanEval, MBPP, our internal eval, leetcode hard, code repair, docstring generation, test generation, and code explanation). Both models at q4_k_m, same hardware (4090), same temperature (0.2), same context limit (8192 tokens). Results across all 8 benchmarks: qwen 2.5 coder 32b: 76% average deepseek v3: 71% average The gap was biggest on code repair (qwen 83%, deepseek 69%) and test generation (qwen 79%, deepseek 68%). Deepseek actually won on docstring generation by 4 points, but that's the only category. Performance: qwen: 18 tok/s average, 23.8gb vram deepseek: 14 tok/s average, 37.2gb vram I know deepseek's MoE architecture is supposed to be more efficient, but at q4 quantization it seems like that advantage disappears. The 671b parameter count sounds impressive until you realize most of it is dead weight during inference. Am I missing something, or is the deepseek hype just because it's "open weights" and people want an alternative to qwen?