It scores almost as high as GLM 5.3 and Kimi 3 on Deep SWE, how is that possible?

I like that it has 512K context. If it's as good as Qwen 3.8 27b when quantized I could use it just for that alone.