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

(twitter.com) by rebekkamikkoa | May 26, 2026 | 0 comments on HN