AI产业跟踪：DeepSeek-V3.1正式发布，持续看好AIagent商业化落地进展

Investment Rating - The investment rating for the industry is "Positive" and is maintained [8] Core Insights - The official release and open-source of DeepSeek-V3.1 occurred on August 21, 2025, featuring 685 billion parameters and an extended context length of 128K, supporting both thinking and non-thinking modes [2][5] - The API pricing structure was adjusted on September 6, 2025, with input prices set at 0.5 CNY per million tokens (cache hit) and 4 CNY per million tokens (cache miss), while output prices are 12 CNY per million tokens [2][5] - The model's capabilities have significantly improved in areas such as Tools use, Coding, and information retrieval, with notable performance in programming and search tasks [10] - The training path focuses on Post-Training optimization and data expansion training, enhancing the model's performance in tool usage and agent tasks [10] - The technical specifications include the use of UE8M0 FP8 Scale parameter precision, which is designed for the next generation of domestic chips, achieving a 75% reduction in traffic compared to traditional FP32 scaling [10] - The pricing strategy has become more competitive, with V3.1 being significantly cheaper than competitors like Claude, enhancing the competitiveness of domestic open-source models [10] - The report emphasizes the ongoing monitoring of DeepSeek's future releases and the expected acceleration of AI application monetization in various verticals such as finance, education, healthcare, and enterprise services [10] Summary by Sections Event Description - DeepSeek-V3.1 was officially released and open-sourced on August 21, 2025, with significant upgrades in parameters and context length [5] Event Commentary - The model's performance has shown substantial improvements, particularly in programming and complex search tasks, indicating a strong competitive edge [10]