春节将至，国产大模型再次迎来上新热潮

Core Insights - The Chinese AI industry is experiencing a surge in new model launches ahead of the Spring Festival, with major players like MiniMax, Zhiyuan, and ByteDance introducing their latest products in what is referred to as the "AI Battle of the Gods" [1][11]. Group 1: MiniMax M2.5 Launch - MiniMax officially launched its new text model, MiniMax M2.5, on February 13, which has significantly improved performance metrics, achieving scores of 80.2% on SWE-Bench Verified and 51.3% on Multi-SWE-Bench, surpassing previous models [4][14]. - The M2.5 model features a "native Spec capability," allowing it to deconstruct architecture and functional planning proactively, mimicking the work of a real architect [4][14]. - The M2.5-lightning version supports an output speed of over 100 transactions per second (TPS), which is approximately double that of mainstream models, with operational costs estimated at $0.3 per million tokens for input and $2.4 per million tokens for output [6][16]. Group 2: Zhiyuan GLM-5 Launch - Zhiyuan AI launched its flagship model GLM-5 on February 12, which saw its stock price surge by 28.68%, reaching a market capitalization of 179.2 billion HKD, an increase of 40 billion HKD from the previous day [7][17]. - The GLM-5 model has expanded its base from 355 billion parameters (activated 32 billion) to 744 billion parameters (activated 40 billion), and its pre-training data has increased from 23 terabytes to 28.5 terabytes, enhancing its general intelligence capabilities [7][17]. - In programming capabilities, GLM-5 aligns with Claude Opus 4.5 and has achieved the highest scores in industry-standard benchmarks, outperforming Gemini 3 Pro [19]. Group 3: ByteDance Seedance 2.0 - ByteDance's AI video generation model, Seedance 2.0, has gained significant attention for its multimodal input capabilities, generating high-quality videos from text and images, which has attracted global interest [1][11]. - The model's rapid development has been noted by industry figures, including Elon Musk and filmmaker Jia Zhangke, highlighting its potential for creative applications [1][11]. Group 4: Future Developments - The industry anticipates further advancements, with Zhiyuan indicating that models will evolve from merely "writing" to "doing tasks," particularly in completing large-scale tasks end-to-end [19]. - Upcoming releases include ByteDance's Doubao Model 2.0, which is expected to excel in multimodal perception and complex task execution, marking a significant upgrade since its initial launch [19][20].