MiniMax-W (00100.HK) has launched its latest model, MiniMaxM3, which boasts state-of-the-art performance in specialized tasks such as programming and agent-based reasoning. The model introduces a new attention architecture, MiniMax Sparse Attention (MSA), enabling it to handle ultra-long contexts of up to 1 million tokens. MiniMaxM3 is a natively multimodal model, capable of processing images and videos, and interacting with computer desktops. It is the first domestic model to integrate these features and remains the only open-source model offering such capabilities, aligning with the standards of leading closed-source models overseas.