[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fi_pVD3pBv-q-TD7saFCS6PFCjcm_EHYboFtunAtJaWA":3},{"code":4,"msg":5,"data":6},200,"操作成功",{"id":7,"title":8,"content":9,"digest":10,"source":10,"coverPath":11,"thumbsCoverPath":12,"isTop":13,"isShow":14,"baseClick":13,"clickCount":15,"createTime":16,"typeId":17,"isNewest":18,"newsInfoTypeRespVo":19,"voiceUrl":22,"voiceSize":23,"taskId":24,"releaseTime":25,"titleEn":26,"contentEn":27,"voiceUrlEn":28,"taskIdEn":29,"voiceSizeEn":30},1264,"DeepSeek发布V3.1模型，迈向Agent时代的关键一步","\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">刚刚，DeepSeek宣布正式推出DeepSeek-V3.1模型，这是一次重大升级，旨在通过创新的混合推理架构、更高的思考效率和更强的智能体能力，为用户提供更高效、更灵活的AI解决方案。本次升级已在官方App、网页端及API平台同步上线，标志着DeepSeek在人工智能领域向“Agent时代”迈出了重要一步。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">DeepSeek-V3.1的核心亮点在于其独特的“混合推理架构”，该架构允许一个模型同时支持“思考模式”和“非思考模式”。用户可以通过官方App或网页端的“深度思考”按钮自由切换模式，实现更智能的交互体验。在思考模式下，模型能显著缩短响应时间，相比前代DeepSeek-R1-0528，DeepSeek-V3.1-Think在输出token减少20%-50%的情况下，保持了相同的任务表现。这一效率提升得益于思维链压缩训练，有效优化了资源消耗。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">在各项评测指标得分基本持平的情况下（AIME 2015: 87.5\u002F88.4, GPQA: 81\u002F80.1, liveCodeBench: 73.3\u002F74.8），R1-0528 与 V3.1-Think 的 token 消耗量对比图：\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002F515d58df7b574fcdadd994e1d7cffda7\u002FAA1KVahv.png\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">此外，DeepSeek-V3.1在智能体能力上实现了质的飞跃。通过Post-Training优化，模型在工具使用和智能体任务中的表现大幅提升。编程智能体方面，在代码修复测评SWE与命令行终端环境下的复杂任务（Terminal-Bench）测试中，DeepSeek-V3.1相比前代模型有显著进步，所需轮数更少。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">表 1：编程智能体测评（SWE 使用内部框架测评，相比开源框架 OpenHands 所需轮数更少；Terminal Bench 使用官方 Terminus 1 framework）：\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002F26b5be2b95754b4faefb2c44674b64fb\u002FAA1KVfJ3.png\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">搜索智能体能力同样得到增强，在需要多步推理的复杂搜索测试（browsecomp）与多学科专家级难题测试（HLE）上，DeepSeek-V3.1性能已大幅领先R1-0528。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">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）：\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002F041d746068f44e5bb9b95001a8ab8fae\u002FAA1KVkJu.png\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">API方面，DeepSeek平台已全面升级：deepseek-chat对应非思考模式，deepseek-reasoner对应思考模式，上下文窗口扩展至128K。同时，API Beta接口支持了strict模式的Function Calling，确保输出符合schema定义（详见官方文档）。新模型还增加了对Anthropic API格式的支持，便于用户将DeepSeek-V3.1集成到Claude Code框架中。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">在开源策略上，DeepSeek-V3.1的Base模型和后训练模型已在Hugging Face和魔搭平台开源。Base模型基于V3进行了840B tokens的外扩训练，后训练模型则针对推理优化。需要注意的是，V3.1采用了UE8M0 FP8 Scale参数精度，并对分词器及chat template进行了调整，建议部署用户参考新版说明文档。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">价格政策方面，DeepSeek将于北京时间2025年9月6日凌晨起调整API接口调用价格，执行新版价格表并取消夜间时段优惠。在9月6日前，所有API服务仍按原价格计费。为满足用户需求，平台已扩容服务资源。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"color: rgb(187, 187, 187);\">【新闻来源】 金融界财经 \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002Fnews\u002Fother\u002Fdeepseek%E5%8F%91%E5%B8%83v3-1%E6%A8%A1%E5%9E%8B-%E8%BF%88%E5%90%91agent%E6%97%B6%E4%BB%A3%E7%9A%84%E5%85%B3%E9%94%AE%E4%B8%80%E6%AD%A5\u002Far-AA1KVhVP?ocid=msedgntphdr&cvid=0ea18e61d8c242ccabdb61117c000caa&ei=33\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(187, 187, 187);\">http:\u002F\u002Fu5a.cn\u002FEx31R\u003C\u002Fa>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(187, 187, 187);\">（本网转发此文章，旨在为读者提供更多的信息资讯，所涉内容不构成投资、消费建议。文章事实如有疑问，请与有关方核实，文章观点非本网观点，仅供读者参考。）\u003C\u002Fspan>\u003C\u002Fp>","","https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002F07100b89caae485ba79b98a7884f3387\u002FAI领域.jpg","https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002Fthumbs\u002F07100b89caae485ba79b98a7884f3387\u002FAI领域.jpg",0,1,253,"2025-08-22 17:07",2,false,{"id":17,"name":20,"enName":21},"芯位视野","Xinwei Vision","https:\u002F\u002Fxinwei-dev-test.oss-cn-shenzhen.aliyuncs.com\u002Fintelligent\u002Faudio%3A20938edc-f2cc-412b-af31-bb94ef64db98%3A0.wav?Expires=1755862028&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=NZsU33JwUtGHbk%2Bs1BzdBJ0djPM%3D",6696974,"20938edc-f2cc-412b-af31-bb94ef64db98","2025-08-22 16:26","DeepSeek launches V3.1 model, a key step towards the Agent era","\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">Just now, DeepSeek announced the official release of the DeepSeek-V3.1 model, which is a major upgrade aimed at providing users with more efficient and flexible AI solutions through an innovative hybrid reasoning architecture, higher thinking efficiency, and stronger agent capabilities. This upgrade has been simultaneously launched on the official app, web version, and API platform, marking a significant step for DeepSeek towards the \"Agent era\" in the field of artificial intelligence.\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The core highlight of DeepSeek-V3.1 is its unique \"hybrid reasoning architecture,\" which allows a single model to support both \"thinking mode\" and \"non-thinking mode.\" Users can freely switch between modes via the \"Deep Thinking\" button on the official app or web version, achieving a more intelligent interaction experience. In thinking mode, the model significantly reduces response time. Compared to the previous DeepSeek-R1-0528, DeepSeek-V3.1-Think maintains the same task performance while reducing output tokens by 20%-50%. This efficiency improvement is due to chain-of-thought compression training, effectively optimizing resource consumption.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In terms of evaluation indicators, the scores are roughly comparable (AIME 2015: 87.5\u002F88.4, GPQA: 81\u002F80.1, liveCodeBench: 73.3\u002F74.8), token consumption comparison between R1-0528 and V3.1-Think:\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002F515d58df7b574fcdadd994e1d7cffda7\u002FAA1KVahv.png\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In addition, DeepSeek-V3.1 has achieved a qualitative leap in agent capabilities. Through post-training optimization, the model's performance in tool usage and agent tasks has significantly improved. In the programming agent aspect, in the code repair evaluation SWE and complex tasks in the command line terminal environment (Terminal-Bench), DeepSeek-V3.1 has made significant progress compared to the previous model, requiring fewer rounds.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">Table 1: Programming Agent Evaluation (SWE uses internal framework for evaluation, requiring fewer rounds than the open-source framework OpenHands; Terminal Bench uses the official Terminus 1 framework):\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002F26b5be2b95754b4faefb2c44674b64fb\u002FAA1KVfJ3.png\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Search agent capabilities have also been enhanced. In complex search tests requiring multi-step reasoning (browsecomp) and multi-disciplinary expert-level problems (HLE), DeepSeek-V3.1's performance has significantly surpassed R1-0528.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">Table 2: Search Agent Evaluation (Test results use commercial search engine API + web filtering + 128K context window; R1-0528 uses internal workflow mode testing; HLE test uses Python and search tools simultaneously):\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002F041d746068f44e5bb9b95001a8ab8fae\u002FAA1KVkJu.png\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Regarding the API, the DeepSeek platform has been fully upgraded: deepseek-chat corresponds to non-thinking mode, deepseek-reasoner corresponds to thinking mode, and the context window has been expanded to 128K. At the same time, the API Beta interface supports strict mode function calling, ensuring that the output conforms to the schema definition (see the official documentation). The new model also adds support for the Anthropic API format, making it easier for users to integrate DeepSeek-V3.1 into the Claude Code framework.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In terms of open-source strategy, the Base model and post-training model of DeepSeek-V3.1 have been open-sourced on Hugging Face and Mota platforms. The Base model was trained on 840B tokens based on V3, and the post-training model is optimized for reasoning. It should be noted that V3.1 uses UE8M0 FP8 Scale parameter precision and has adjusted the tokenizer and chat template. It is recommended that deployment users refer to the new documentation.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Regarding pricing policy, DeepSeek will adjust the API call prices starting at midnight Beijing Time on September 6, 2025, implementing the new price list and canceling the night-time discount. Before September 6, all API services will still be billed at the original price. To meet user needs, the platform has expanded service resources.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"color: rgb(187, 187, 187);\">【News Source】 FinanceNet \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002Fnews\u002Fother\u002Fdeepseek%E5%8F%91%E5%B8%83v3-1%E6%A8%A1%E5%9E%8B-%E8%BF%88%E5%90%91agent%E6%97%B6%E4%BB%A3%E7%9A%84%E5%85%B3%E9%94%AE%E4%B8%80%E6%AD%A5\u002Far-AA1KVhVP?ocid=msedgntphdr&cvid=0ea18e61d8c242ccabdb61117c000caa&ei=33\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(187, 187, 187);\">http:\u002F\u002Fu5a.cn\u002FEx31R\u003C\u002Fa>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(187, 187, 187);\">（This article is forwarded by this site to provide readers with more information and news. The content involved does not constitute investment or consumption advice. If there are any questions about the facts of the article, please verify with the relevant parties. The views of the article are not the views of this site and are for reference only.)\u003C\u002Fspan>\u003C\u002Fp>","https:\u002F\u002Fxinwei-dev-test.oss-cn-shenzhen.aliyuncs.com\u002Fintelligent\u002Faudio%3Ad13676b2-9e83-4b46-ba5e-b0233a47c106%3A0.wav?Expires=1774838489&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=VP6R0mlIK02FEFBgcCirt9Or7hM%3D","d13676b2-9e83-4b46-ba5e-b0233a47c106",8033356]