As we've come to expect from a DeepSeek release， DeepSeek V4 comes with more flashy ML systems optimizations. This time？ MegaMoE， a 1400 line fused CUDA kernel that computes the entire MoE forward pass. Let's see how it works （1/4） 🧵