蚂蚁发布万亿参数旗舰模型Ling-1T并开源

Core Insights - Ant Group has launched its flagship general language model Ling-1T, which is the largest and most capable model in the Ling 2.0 series to date [1][2] - Ling-1T has demonstrated state-of-the-art (SOTA) performance in complex reasoning benchmarks, achieving a balance between efficient thinking and precise reasoning [1] - The model excels in various high-difficulty benchmark tests, including code generation and mathematical reasoning, outperforming many open-source models [1] Model Performance - In the AIME 25 competition mathematics benchmark, Ling-1T achieved a 70.42% accuracy rate with an average consumption of over 4000 tokens, surpassing Gemini-2.5-Pro, which had a 70.10% accuracy rate with over 5000 tokens [1] - Ling-1T's ability to achieve higher accuracy with fewer tokens highlights its advantages in reasoning precision and thinking efficiency [1] Technical Specifications - Ling-1T is built on the Ling 2.0 architecture and has been pre-trained on over 20 trillion tokens of high-quality, high-density reasoning data [2] - The model supports a maximum context window of 128K tokens and utilizes an evolutionary thinking chain (Evo-CoT) approach to enhance its efficient thinking and precise reasoning capabilities [2]