Concept稀疏注意力
MiniMax Sparse Attention
A sparse attention technique proposed by MiniMax in which each token attends only to a subset of positions (a local window plus a small number of global tokens), reducing computation from quadratic to near-linear and substantially lowering the cost of long-context inference.
Timeline (last 90 days)
Oct 19
MiniMax M3 is built on MiniMax Sparse Attention, making the operating cost of a 1 million token context far lower than the previous generation.
Unverified50%