Qwen 4 Max Performance Results on KernelBench-CUDA

x.com2 min read

Results for Qwen 4 Max on KernelBench-CUDA reveal that the model wrote three hand-written CUDA kernels for an RTX PRO 6000 over multi-hour sessions. The tests covered Native Sparse Attention, Fused MoE, and MegaQwen Decode, demonstrating deep debugging and profiling capabilities alongside notable misdiagnoses and deadlocks. Open article →