MiniMax深夜开源首个推理模型M1，这次是真的卷到DeepSeek了。

Core Viewpoint - The article discusses the recent release of MiniMax's first inference model, MiniMax M1, which is claimed to have context capabilities comparable to the leading model, Gemini 2.5 Pro [2][10]. Group 1: Model Performance - MiniMax M1 has shown competitive performance in various benchmarks, particularly excelling in the MRCR (Multi-Round Co-reference Resolution) task, achieving an accuracy of 62.8%, which is on par with Gemini 2.5 Pro [3][8]. - The model's architecture includes 456 billion parameters with a MoE (Mixture of Experts) structure, allowing it to handle a maximum context length of 1 million words, significantly surpassing DeepSeek-R1's capabilities [10][12]. - The Lightning Attention mechanism used in MiniMax M1 allows for linear growth in time and space complexity with increasing sequence length, making it more efficient than traditional transformers [8][9]. Group 2: Benchmark Comparisons - In the AIME 2024 logic and mathematics tasks, MiniMax M1 performed adequately, with some tasks showing strong results while others were average [3]. - The MRCR task, which tests a model's ability to understand and differentiate between multiple conversation threads, is highlighted as a significant challenge that MiniMax M1 has managed to tackle effectively [6][8]. Group 3: User Experience and Applications - Users have reported impressive experiences with MiniMax M1, including its ability to accurately translate complex documents and maintain context over long interactions [14][22]. - The model's capabilities extend to creative applications, such as generating narrative content and engaging in interactive storytelling, showcasing its versatility [31][33]. Group 4: Future Expectations - There is anticipation for further developments from MiniMax, particularly in video models and other innovative applications, as the company continues to push the boundaries of AI technology [42][46].