DeepSeek V4.1 Flash reportedly ranks first of 23 on MegaQwen decode

A post reports that DeepSeek V4.1 Flash (1017) passed all four handwritten CUDA kernels on an RTX Pro 6000 and ranked first of 23 on MegaQwen decode, reaching 8.12% of roofline against Sonnet 5.5’s 8.00%. Its decode KV cache used FP8 despite the problem specifying BF16. The post says its trace treated the 0.08 error tolerance as “real slack”; cosine similarity to the reference was 0.99994 at 32k context. It calls the result a precision trade-off, not a free win.