MiniMax dropped a new attention architecture. [N]

It contains something interesting about context windows. They’re natively scaling to 1M tokens with MiniMax Sparse Attention (MSA) , bypassing standard quadratic complexity by completely restructuring the memory access patterns at the operator level. Instead of relying on typical sparse approximations that degrade recall, MSA utilizes a clean " KV outer gather Q " approach. By treating KV blocks as the outer loop to aggregate hit queries, hardware memory reads remain strictly contiguous, and each block is f