8

MiniMax teased M3 Sparse Attention: 9.7x prefilling, 15.6x decoding at 1M