MiniMaxAI/MiniMax-M3.1-Flash (huggingface.co)

MiniMaxAI has unveiled MiniMax-M3.1-Flash, a cutting-edge multimodal mixture-of-experts model designed for tasks in coding, agentic workflows, and cybersecurity. This new model enhances its predecessor, MiniMax-M3, by significantly improving task performance and inference efficiency through innovations like sparse attention, quantization-aware training, and DSpark speculative decoding. Notable performance metrics include scores of 76.2 on the BankerToolBench and 82.2 on CyberGym, showcasing its remarkable capabilities in various benchmark evaluations. The significance of MiniMax-M3.1-Flash lies in its architecture, which supports a massive context window of 1 million tokens and can process text, images, and videos simultaneously, making it highly versatile for developers and researchers in the AI/ML field. Its design involves approximately 428 billion parameters, with only 23 billion activated per token, leveraging MiniMax Sparse Attention across the model for improved efficiency. These advancements not only enhance computational performance but also reduce memory usage, indicating a transformative step for deploying large-scale AI applications effectively.