MiniMax teased M3 Sparse Attention: 9.7x prefilling, 15.6x decoding at 1M #1 Post by rebekkamikkoa » Tue, May 26, 2026, 9:09 PM UTC MiniMax teased M3 Sparse Attention: 9.7x prefilling, 15.6x decoding at 1Mtwitter.com