大模型“上海队”进入丰产阶段（神州看点） 生成的“猫跳水”视频一周获三亿播放量

Core Insights - MiniMax, a Shanghai-based AI company, has launched the world's first open-source large-scale hybrid architecture inference model, MiniMax-M1, which ranks second globally among open-source models [1] - The company has also released video generation model Hailuo 02, which achieved 300 million views within a week of its release on social media [1][6] - MiniMax distinguishes itself by not following mainstream dense architectures and traditional attention mechanisms, focusing instead on AGI since before the rise of ChatGPT [1][8] Performance and Cost Efficiency - The competition in large models is shifting from mere parameter scale to efficiency, cost, and overall implementation capabilities [2] - M1 supports an impressive context input of 1 million tokens, comparable to Google's latest closed-source model Gemini 2.5 Pro, while its reinforcement learning phase cost only $535,000 [2] - Hailuo 02 directly competes with Google's third-generation video generation model Veo3, showcasing superior performance in generating coherent and logical video sequences [3] Innovation in AI Video Generation - Hailuo 02 has pioneered a new category of AI video called "Animal Olympics" [4] - The development of Hailuo 02 involved collaboration with a diverse team of directors, screenwriters, and artists to ensure high-quality output [5] - High-quality data, innovative algorithms, and meticulous training processes are cited as key factors in the success of Hailuo 02 [6] Strategic Positioning - MiniMax remains one of the few startups still committed to foundational model research amidst a trend of major companies reducing their efforts in this area [7] - The company is exploring "sparse activation" MoE architecture to reduce computational costs, diverging from the prevalent dense architecture approach [8] - MiniMax aims to stay competitive in the long-term race of large model development, collaborating with other major players in Shanghai's AI ecosystem [9]