MiniMax teased M3 Sparse Attention: 9.7x prefilling, 15.6x decoding at 1M twitter.com 6 points by rebekkamikkoa 5 hours ago