MiniMax unveils M3 AI with faster response speeds

MiniMax has revealed details about its upcoming M3 series of AI models, highlighting a new sparse attention mechanism designed to significantly boost response speed. This innovation promises up to a 15.6 times faster decoding speed, making the deployment of AI agents with ultra-long context economically viable. The company's technical report on the M2 series also sheds light on its engineering advancements, including a sparse Mixture-of-Experts architecture and strategies to overcome the computational challenges of standard attention mechanisms. MiniMax's approach aims to balance high-speed processing with robust reasoning capabilities, addressing a key bottleneck in current large language models.