Results for Qwen 4 Max on KernelBench-Mega

x.com2 min read

Qwen 4 Max achieved 16.3x the optimized PyTorch baseline on the KernelBench-Mega benchmark for Triton. The model built its own grid barrier using atomic counters to handle decoding tasks, though an attempt to remove barriers entirely resulted in compilation failures and deadlocks. Open article →