智通财经APP获悉，8月21日，DeepSeek正式发布 DeepSeek-V3.1。本次升级包含主要变化有：混合推理架构(一个模型同时支持思考模式与非思考模式)； 更高的思考效率(相比 DeepSeek-R1-0528，DeepSeek-V3.1-Think 能在更短时间内给出答案)；更强的 Agent 能力(通过 Post-Training 优化，新模型在工具使 用与智能体任务中的表现有较大提升)。 表 2：搜索智能体测评(测试结果调用商用搜索引擎 API+网页过滤+128K context window；R1-0528 使用内部 workflow 模式测试；HLE 测试同时使用 python 与 search 工具) | Benchmarks | DeepSeek-V3.1 | DeepSeek- | | --- | --- | --- | | | | R1-0528 | | Browsecomp | 30.0 | 8.9 | | Browsecomp_zh | 49.2 | 35.7 | | HLE | 29.8 | 24.8 | | xbench-DeepSearch | 71.2 | 55.0 | ...

8月21日，《财富》杂志发布了2025年中国科技50强榜单，华为投资控股有限公司、深度求索（DeepSeek）、宁德时代新能源科技股份有限公司等多家科 技企业入选。该榜单涵盖人工智能、新能源、生物医药、高端制造等多个前沿领域。 据《财富》介绍，在技术迭代与市场需求的交汇中，中国公司们正在以强大的执行力推动科技落地。在大模型领域，它们避开空泛的概念炒作，聚焦金 融、医疗等垂直领域的模型优化，提高效率；它们通过优化关节灵活性与续航能力，让机器替代人类进入高危、高强度作业场景；它们研发更高效、清洁 的能源，让人类与自然更和谐相处。 除上述公司外，阿里巴巴、腾讯、比亚迪、字节跳动等知名科技企业也跻身该榜单。值得注意的是，不少新兴科技公司如宇树科技、稀宇科技、玻色量子 等同样凭借在细分领域的创新表现获得认可。 | 北京字节跳动科技有限公司 | | --- | | 杭州宇树科技股份有限公司 | | 大疆创新 | | 石药控股集团有限公司 | | 联想集团有限公司 | | 京东方科技集团股份有限公司 | | 上海稀宇科技有限公司 | | 大族激光科技产业集团股份有限公司 | | 杭州云深处科技有限公司 | | 顺丰科技有限 ...

Core Insights - DeepSeek has officially released the upgraded model DeepSeek-V3.1, which features a hybrid reasoning architecture that supports both thinking and non-thinking modes [1] - The new model, DeepSeek-V3.1-Think, demonstrates improved thinking efficiency, providing answers in a shorter time compared to its predecessor DeepSeek-R1-0528 [1] - Enhanced agent capabilities have been achieved through post-training optimization, significantly improving performance in tool usage and agent tasks [1] Pricing Adjustments - Starting from September 6, 2025, the pricing for API calls on the DeepSeek open platform will be adjusted according to a new price list, with the cancellation of night-time discounts [2] - Until September 6, 2025, all API services will continue to be billed under the original pricing policy [4]

Core Insights - DeepSeek announced the release of DeepSeek-V3.1 and will adjust the API pricing effective September 6, 2025 [1][3] - The new pricing structure includes input prices of 0.5 CNY per million tokens for cache hits and 4 CNY per million tokens for cache misses, with output prices set at 12 CNY per million tokens [1] Group 1: Upgrade Features - The V3.1 upgrade introduces a hybrid reasoning architecture that supports both thinking and non-thinking modes within a single model [3] - Enhanced thinking efficiency allows DeepSeek-V3.1-Think to provide answers in a shorter time compared to its predecessor, DeepSeek-R1-0528 [3] - Improved agent capabilities through post-training optimization significantly enhance the model's performance in tool usage and agent tasks [3] Group 2: User Experience - The official app and web model have been upgraded to DeepSeek-V3.1, allowing users to switch freely between thinking and non-thinking modes via a "deep thinking" button [3]

Core Insights - DeepSeek has officially released DeepSeek-V3.1, which includes significant upgrades in its architecture and performance [1] Group 1: Key Features of DeepSeek-V3.1 - Hybrid reasoning architecture: The model supports both thinking and non-thinking modes simultaneously [1] - Enhanced thinking efficiency: DeepSeek-V3.1-Think can provide answers in a shorter time compared to DeepSeek-R1-0528 [1] - Improved agent capabilities: The new model shows significant improvements in tool usage and agent tasks through post-training optimization [1]

Group 1 - DeepSeek officially released DeepSeek-V3.1, featuring a hybrid reasoning architecture that supports both thinking and non-thinking modes [1] - The new version, DeepSeek-V3.1-Think, offers higher thinking efficiency, providing answers in a shorter time compared to DeepSeek-R1-0528 [1] - Enhanced agent capabilities have been achieved through post-training optimization, significantly improving performance in tool usage and intelligent tasks [1] Group 2 - Starting from September 6, 2025, the pricing for API calls on the DeepSeek open platform will be adjusted, with input costs set at 0.5 to 4 yuan per million tokens (cache hit) and 4 yuan per million tokens (cache miss), while output costs will be 12 yuan per million tokens [1]

本次升级包含以下主要变化：混合推理架构：一个模型同时支持思考模式与非思考模式；更高的思考效 率：相比DeepSeek-R1-0528，DeepSeek-V3.1-Think能在更短时间内给出答案；更强的Agent能力：通过 Post-Training优化，新模型在工具使用与智能体任务中的表现有较大提升。 官方App与网页端模型已同步升级为DeepSeek-V3.1。用户可以通过"深度思考"按钮，实现思考模式与非 思考模式的自由切换。 （文章来源：第一财经） 据DeepSeek官方公众号消息，DeepSeek-V3.1正式发布。 ...

Core Insights - The "Fortune China Top 50 Technology Companies" list was released, featuring companies like Huawei, DeepSeek, and Yushu Technology [1] Group 1: DeepSeek - DeepSeek is recognized as a leading AI large model product in China, with its DeepSeek-R1 model scoring 88.5 on the MMLU benchmark test, which is lower than OpenAI's GPT-4 (92.0) and Google's Gemini Pro (90.0), but higher than Meta's Llama 3 (82.0) and Anthropic's Claude 2 (85.1) [1] - DeepSeek ranks among the top 10 globally in terms of open-source large model downloads, indicating strong market presence [1] - As of June 2025, DeepSeek is projected to have 163 million monthly active users, making it the leading application in AI-generated content globally [1] Group 2: Yushu Technology - In 2024, Yushu Technology achieved global sales of 18,000 quadruped robots, capturing a 23% market share, ranking second only to Boston Dynamics [1] - Yushu Technology was awarded the WIPO 2025 Global Award, distinguishing it as the only representative from China among 780 applicants from 95 countries and regions [1] - The company's success is attributed to innovations in robotic motion control, high-performance joint motors, and real-time systems, along with a comprehensive global intellectual property strategy [1]

Core Insights - DeepSeek has released an updated version of its model, V3.1, which shows significant improvements in context length and user interaction, although it is not the highly anticipated R2 model [2][4][14] - The model now supports a context length of 128K, enhancing its ability to handle longer texts and improving its programming capabilities [5][10] - The update merges the functionalities of V3 and R1, leading to reduced deployment costs and improved efficiency [13][25] Group 1: Model Improvements - The new V3.1 model has a parameter count of 685 billion, showing only a slight increase from the previous version, V3, which had 671 billion parameters [7] - User experience has been enhanced with more natural language responses and the use of tables for information presentation [8][10] - The programming capabilities of V3.1 have been validated through tests, achieving a score of 71.6% in multi-language programming, outperforming Claude 4 Opus [10] Group 2: Market Context - The release of V3.1 comes seven months after the launch of R1, during which time other major companies have also released new models, using R1 as a benchmark [3][16] - Despite the improvements in V3.1, the industry is still eagerly awaiting the release of the R2 model, which has not been announced [4][20] - The competitive landscape includes companies like Alibaba and ByteDance, which have launched models that claim to surpass DeepSeek R1 in various metrics [17][19] Group 3: Future Outlook - There are indications that the merging of V3 and R1 may be a preparatory step for the release of a multi-modal model [25] - Industry insiders suggest that the focus will shift towards innovations in economic viability and usability for future models [24] - The absence of the R2 model in the current update has heightened expectations for its eventual release, with speculation that it may not arrive until later [21][22]

Core Viewpoint - The recent upgrade of DeepSeek to version 3.1 has shown significant improvements in context length and user interaction, while also merging features from previous models to reduce deployment costs [1][11][12]. Group 1: Model Improvements - DeepSeek V3.1 now supports a context length of 128K, enhancing its ability to handle longer texts [4]. - The model's parameter count increased slightly from 671 billion to 685 billion, but the user experience has improved noticeably [5]. - The model's programming capabilities have been highlighted, achieving a score of 71.6% in multi-language programming tests, outperforming Claude 4 Opus [7]. Group 2: Economic Efficiency - The merger of V3 and R1 models allows for reduced deployment costs, requiring only 60 GPUs instead of the previous 120 [12]. - Developers noted that the performance could improve by 3-4 times with the new model due to increased cache size [12]. - The open-source release of DeepSeek V3.1-Base on Huggingface indicates a move towards greater accessibility and collaboration in the AI community [13]. Group 3: Market Context - The AI industry is closely watching the developments of DeepSeek, especially in light of the absence of the anticipated R2 model [19]. - Competitors like OpenAI, Google, and Alibaba have released new models, using R1 as a benchmark for their advancements [1][15]. - The market is eager for DeepSeek's next steps, particularly regarding the potential release of a multi-modal model following the V3.1 update [23].