is live. can frontier models write fast triton/cuda/cutlass/cute-dsl/ptx without cheating? i add kernelbench-hard to the existing kernelbench-v3 (which was built on top of KernelBench from @anneouyang et al.) hard mode has: fp8 gemm, topk, sonic MoE fwd, KimiDeltaAttention, paged attention decode, kahan softmax, w4a16 gemm. all of these require deep understanding of the sm120 architecture (benchmarking happens on my local rtx pro 6000 blackwell setup). gpt 5.5 xhigh and claude opus 4.7 max clear took the W here, but i was honestly surprised with kimi-k2.6 and deepseek v4 pro. this is just the first public iteration. im open to constructive criticism (dm works)