In my benchmark Deepseek-v4-flash did much better than Qwen 3.8 27B at reverse engineering.

https://alexander-hanel.github.io/StressingLLMs/