Core Insights - DeepSeek has released an upgraded version, DeepSeek-V3.1-Terminus, which shows significant performance improvements over its predecessor [1][2] - The new version achieved a score of 21.7 in the "Humanitys Last Exam" test, a 36% increase from the previous score of 15.9, and a 28.3% improvement in the "BrowseComp" test [1][2] Company Developments - The API pricing remains unchanged, with input costs at 0.5 yuan per million tokens and output costs at 12 yuan [1] - The update addresses user feedback by enhancing language consistency and improving the performance of Code Agent and Search Agent [2] Industry Context - The rapid updates and bug fixes in the AI model are crucial for retaining developers in a highly competitive open-source model landscape [2] - Analysts highlight DeepSeek's advantages, including high performance, low cost, and innovative algorithms, positioning it competitively against global counterparts [2] - The demand for AI computing power is expected to grow, benefiting various segments of the computing industry, including AI chips and server systems [2]

Core Insights - The new version DeepSeek-V3.1-Terminus has been launched, featuring both "Thinking Mode" and "Non-Thinking Mode" with support for 128K long context [1] Group 1: Model Upgrades - The deepseek-chat and deepseek-reasoner models have been unified and upgraded to DeepSeek-V3.1-Terminus, with deepseek-chat corresponding to Non-Thinking Mode and deepseek-reasoner to Thinking Mode [1] - Key optimizations include improved language consistency, significantly alleviating issues with mixed Chinese and English as well as abnormal characters, resulting in more standardized outputs [1] - The Agent capabilities have been further enhanced, particularly the execution performance of Code Agent and Search Agent [1] Group 2: Output Length and Pricing - In terms of output length, Non-Thinking Mode supports a default of 4K, with a maximum of 8K, while Thinking Mode has a default of 32K and can be expanded up to 64K, catering to different generation length requirements [1] - Pricing for the new model is set at 0.5 yuan for cache hits and 4 yuan for cache misses per million tokens input, with an output pricing of 12 yuan per million tokens, providing developers with a cost-effective AI large model service [1]

Core Insights - DeepSeek has released an updated version, DeepSeek-V3.1-Terminus, which shows significant performance improvements over its predecessor [1][3] - The new version's scores in various benchmarks have increased, with "Humanity's Last Exam" score rising from 15.9 to 21.7, a more than 36% improvement [1][2] - The update addresses user feedback, enhancing language consistency and agent capabilities, which are crucial in the competitive landscape of open-source models [3] Company Analysis - DeepSeek demonstrates high performance, low cost, and effective distillation, showcasing China's competitive strength in AI technology innovation [4] - The rapid updates and bug fixes are essential for retaining developers in a highly competitive environment [3] - The release of new computing architectures is expected to drive demand for AI computing power, benefiting various segments of the computing industry [4]

Group 1 - Aerospace ETF Tianhong (159241) closed at a gain of 0.35% as of September 22, 2025, marking two consecutive days of increase with a turnover rate of 27.34% and a transaction volume of 120 million yuan, indicating active market trading [2] - The latest scale of Aerospace ETF Tianhong reached 491 million yuan with a total of 429 million shares, achieving a new high in nearly one month [2] - The net inflow of funds into Aerospace ETF Tianhong was 47.7466 million yuan, with a total of 51.0593 million yuan accumulated over the last five trading days [2] Group 2 - Aerospace ETF Tianhong (159241) serves as an efficient tool to capture core military aerospace opportunities, tracking the National Securities Aerospace Index, with a military attribute of 97.86%, making it the highest military content index in the market [3] - The weight of aerospace equipment in the ETF is as high as 66.8%, significantly exceeding that of the CSI Military and CSI National Defense indices [3] Group 3 - The recent successful launch and recovery training of the Navy's third type of carrier-based aircraft on the Fujian ship marks a significant breakthrough in China's aircraft carrier development, contributing to the transformation of the Navy [4] - The Navy's carrier-based aviation has accelerated its development, achieving a significant leap from single aircraft to a system, from land-based to carrier-based operations, and from conventional takeoff to electromagnetic catapult [4] Group 4 - Zhonghang Securities assesses that the military industry sector is currently in a state with ample upward potential and a solid bottom, with short-term volatility risks in rapidly rising sub-sectors and individual stocks, but a relatively low likelihood of structural deep adjustments [5] - The military industry is expected to perform better in the second half of the year compared to the first half, with emerging themes such as low-altitude economy, commercial aerospace, deep-sea technology, large aircraft, and military intelligence likely to remain active [5]

Core Insights - DeepSeek announced the update of its model to DeepSeek-V3.1-Terminus, enhancing its existing capabilities while addressing user feedback [1] Improvements - The update focuses on two main areas: - Language consistency, which alleviates issues related to mixed Chinese and English text and occasional abnormal characters [1] - Enhanced performance of intelligent agents, specifically the Code Agent and Search Agent [1]

Core Insights - DeepSeek has released an updated model, DeepSeek-V3.1-Terminus, which improves upon the previous version by enhancing language consistency and fixing bugs related to unexpected character outputs [1][7][20] - The model has been open-sourced, allowing broader access and potential community contributions [1][7] Performance Improvements - Benchmark tests show that DeepSeek-V3.1-Terminus has achieved performance improvements ranging from 0.2% to 36.5% compared to DeepSeek-V3.1, with notable enhancements in the Human's Last Exam (HLE) test, which assesses high-level knowledge and reasoning capabilities [3][5] - In non-Agent evaluations, the model's performance in MMLU-Pro improved from 84.8 to 85.0, and in HLE, it increased from 15.9 to 21.7 [5] Bug Fixes - The previous version had a significant bug where the model would output random characters, which has been resolved in the new version [7][8] - Additionally, issues with multilingual outputs, particularly in translating minor languages, have also been addressed, resulting in more coherent translations [9][10] Enhanced Capabilities - The model demonstrates improved programming and search capabilities, successfully simulating physical effects in programming tasks and providing comprehensive recommendations for plant care based on specific criteria [13][17] - The model's ability to cross-verify information and present it in a readable format has also been highlighted as a significant improvement [17] Future Outlook - The name "Terminus" suggests that this version may represent the culmination of the current technological path for DeepSeek, although future updates, including an Agent model, are anticipated by the end of the year [20][21]

Core Insights - DeepSeek has officially upgraded its model to DeepSeek-V3.1-Terminus, enhancing performance based on user feedback, particularly in language consistency and agent capabilities [1][2] - The new model shows improved stability in output, with benchmark results indicating significant performance gains across various assessments [1] - The release of DeepSeek-V3.1 is seen as a breakthrough for domestic large models and chip ecosystems, addressing compatibility issues with NVIDIA's FP8 standard [2][3] Model Performance - The benchmark results for DeepSeek-V3.1-Terminus compared to its predecessor are as follows: - MMLU-Pro: 85.0 (up from 84.8) - GPQA-Diamond: 80.7 (up from 80.1) - Humanity's Last Exam: 21.7 (up from 15.9) - BrowseComp: 38.5 (up from 30.0) - SimpleQA: 96.8 (up from 93.4) - SWE Verified: 68.4 (up from 66.0) [1] Industry Impact - The launch of DeepSeek V3.1 has significantly boosted the domestic computing industry, with expectations for increased applications of domestic AI chips in training and inference [3][4] - The success of DeepSeek is viewed as a victory for open-source models, prompting other Chinese companies to adopt similar open-source strategies [3] - The AI computing demand is projected to grow, benefiting various segments of the computing supply chain, including AI chips and servers [4] Related Developments - DeepSeek's research paper on the R1 reasoning model has been featured on the cover of the prestigious journal Nature, marking a significant achievement in the field [2] - Other companies in the industry, such as Baidu and Alibaba, are also advancing their models, with Baidu's Wenxin model showing a 34.8% improvement in factual accuracy [6] and Alibaba launching its Qwen3-Max-Preview model [6]

Monetary Policy - The central bank announced a 240.5 billion yuan reverse repurchase operation for 7-day terms and a 300 billion yuan operation for 14-day terms, marking the first 14-day operation in 8 months [2][4]. Industry Development - The Ministry of Industry and Information Technology and the National Development and Reform Commission issued guidelines for the high-quality development of industrial parks, emphasizing the acceleration of green facility construction and the development of new energy infrastructure [2][4]. Sports and Health - The National Sports Administration released guidelines aimed at enhancing public service systems for national fitness and promoting the integration of sports and health [4]. Capital Market Insights - The China Securities Regulatory Commission reported that over 90% of new listed companies in recent years are technology firms, with the market capitalization of the technology sector now exceeding 25% of the total A-share market [7][8]. Semiconductor Industry - Domestic chip stocks surged, with companies like Haiguang Information and Chip Origin seeing gains over 10%. The rapid growth in demand for computing power is expected to drive the GPU and semiconductor industries [10]. Company Performance - Changchuan Technology projected a net profit of 827 million to 877 million yuan for the first three quarters of 2025, representing a year-on-year increase of 131.39% to 145.38% due to strong demand in the semiconductor market [11]. Investment Activities - Zhangjiang Hi-Tech announced an investment of 22.345 million yuan in Shanghai Microelectronics, holding a 10.779% stake [13]. - Huadian International plans to increase its registered capital in a joint investment company by 5 billion yuan, maintaining a 12% stake [15]. - Huahai Pharmaceutical received approval for a new drug, which is expected to have significant therapeutic effects [16]. Market Transactions - Institutional investors bought a net 4.54 billion yuan of Jucheng shares, accounting for 23.62% of the total trading volume [21]. - One institutional investor purchased 2.73 billion yuan of Chip Origin shares, which is part of a strategic acquisition plan to enhance its competitive edge in the RISC-V CPU market [22].

Core Insights - DeepSeek has officially upgraded its model to DeepSeek-V3.1-Terminus, enhancing performance based on user feedback and improving language consistency and agent capabilities [1][2] - The new model shows improved stability in output, with benchmark results indicating performance increases in various assessments compared to the previous version [1] - The release of DeepSeek V3.1 is seen as a significant breakthrough for domestic large models and chip ecosystems, reducing reliance on NVIDIA standards and promoting domestic computing power autonomy [2][3] Model Performance - The benchmark results for DeepSeek-V3.1-Terminus show improvements in several areas, including: - MMLU-Pro: 84.8 to 85.0 - Humanity's Last Exam: 15.9 to 21.7 - SimpleQA: 93.4 to 96.8 - BrowseComp: 30.0 to 38.5 [1] - The model's agent capabilities have significantly improved, which is expected to enhance commercial applications of AI agents [3] Industry Impact - The launch of DeepSeek V3.1 has led to a surge in the domestic computing industry, with increased demand for AI chips and related infrastructure [3][4] - The success of DeepSeek is viewed as a victory for open-source models, prompting other Chinese companies to adopt similar open-source strategies [3] - The AI computing demand is projected to grow, benefiting various segments of the computing industry, including AI chips, servers, and related technologies [4] Related Companies - Baidu has released its Wenxin model X1.1, showing significant improvements in performance metrics compared to previous versions and competing models [6] - Alibaba's Tongyi Qianwen has launched the Qwen3-Max-Preview model, marking advancements in the domestic large model sector [6] - SenseTime's new interactive platform integrates with Xiaomi AI glasses, showcasing the application of AI in real-world scenarios [7] - ZTE has introduced several products focused on AI and intelligent computing, facilitating the deployment of DeepSeek models across various industries [7]

Core Insights - The update maintains the original capabilities of the model while addressing user feedback issues [1] Group 1 - DeepSeek has been updated to version DeepSeek-V3.1-Terminus [1] - Improvements include language consistency, alleviating mixed Chinese and English usage, and occasional abnormal characters [1] - The performance of Code Agent and Search Agent has been further optimized [1]