Chinese AI lab MiniMax claims its next-gen M3 model achieves 15.6x faster decoding using novel sparse attention technology.
