腾讯，大动作！价格为DeepSeek-R1的1/4

Core Viewpoint - Tencent has launched the official version of its self-developed deep thinking model, Mix Yuan T1, indicating a significant advancement in China's AI capabilities aimed at overseas developers [1][2]. Group 1: Model Performance - Mix Yuan T1 has shown improved reasoning capabilities through large-scale reinforcement learning, outperforming DeepSeek-R1 in common benchmark tests like MMLU-PRO [1][2]. - In the DROP F1 test, Mix Yuan T1 also scored higher than DeepSeek-R1 and OpenAI-O1, although it lagged behind in mathematical and coding abilities [2]. - Overall, Mix Yuan T1's performance is now at the level of leading reasoning models in the industry, although Tencent has not disclosed the model's parameter scale [2]. Group 2: Model Architecture - Mix Yuan T1 is built on the Tencent Quick Thinking model, Mix Yuan Turbo S, which emphasizes fast responses and excels in processing long texts [3]. - The architecture of Mix Yuan Turbo S innovatively combines Hybrid-Mamba-Transformer, allowing it to efficiently handle long sequences while capturing complex contexts [3]. - This architecture reduces the computational complexity of traditional Transformer structures, significantly lowering training and inference costs, achieving a token output speed of up to 80 tokens/s [3]. Group 3: Pricing Strategy - The input price for Mix Yuan T1 is set at 1 yuan per million tokens, while the output price is 4 yuan per million tokens, which is competitive compared to DeepSeek-R1's pricing [4]. - During standard hours, Mix Yuan T1's pricing is only one-fourth of DeepSeek-R1's, making it an attractive option for developers [4].