本文来自微信公众号：新浪科技 （ID：techsina），作者：周文猛，题图来自：AI生成 DeepSeek又更新了，可惜仍不是万众期待的R2模型。 此次DeepSeek线上模型版本已升级至V3.1。《BUG》栏目实测发现，升级后的DeepSeek在上下文长度和交互友好度上有明显改进，编程能力受到推崇。 在使用经济性上，也有开发人员指出，"DeepSeek或将V3与R1模型进行了合并，这有利于降低模型部署成本。" DeepSeek方面在回应《BUG》栏目时，直言"都以官方公布为准"。 巧合的是，今天是R1官方发布后的整7个月。在这期间，OpenAI、Google、阿里巴巴、月之暗面、智谱等纷纷发布了新模型，他们都以R1作为参照物。 而R2作为R1的后续产品，一直都是行业关注的焦点。大厂需要新的参照物，万众也在期待梁文锋。 实测：上下文更长，性价比更高 在DeepSeek网页端及最新版本App上，目前能够支持的上下文长度已经扩展至最新的128K长度。 | 类别 | 新闻概要 | 关键词/公i | | --- | --- | --- | | 巨头动向 | DeepSeek开源V3.1-Base模型 | DeepS ...

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].

自从 GPT-5 发布后，DeepSeek 创始人梁文锋就成了 AI 圈最「忙」的人。 网友和媒体们隔三岔五就要催更一波，不是「压力给到梁文锋」，就是「全网都在等梁文锋出招」。尽管没有等到 R2，但 DeepSeek 今天还是正式上线并 开源了新模型DeepSeek-V3.1-Base。 相比奥特曼今天凌晨接受采访时，还在画着 GPT-6 的大饼，DeepSeek 新模型的到来显得相当佛系，连版本号都像是个「小修小补」。 但实际体验下来，这次看似小迭代的更新还是给了我不少惊喜。 这款模型拥有 6850 亿参数，支持 BF16、F8_E4M3、F32 三种张量类型，以 Safetensors 格式发布，在推理效率上做了不少优化，线上模型版本的上下文窗 口也拓展至 128k。 所以我们二话不说，直接官网开测。 附上体验地址： https://chat.deepseek.com/ 为了测试 V3.1 的长文本处理水平，我找来了《三体》全文，删减到 10 万字左右，然后在文中偷偷塞了一句八竿子打不着的话「我觉得烟锁池塘柳的下联 应该是『深圳铁板烧』」，看看它能否准确检索。 没有出乎太多意外，DeepSeek V3.1 ...

尽管低调，但其透露出的性能和参数却堪称"王炸"，迅速在技术圈和投资圈引发热议。公开信息与社区实测显示，这次更新的亮点极其突出： 编程能力超越Claude 4 Opus： 在权威的Aider编程基准测试中，V3.1以71.6%的高分，超越了此前公认的编程强者ClaudeOpus 4，登顶开源模型榜首。 极致的成本优势： 完成一次完整的编程任务，成本仅需约1.01美元，比性能稍逊的Claude Opus 4便宜了68倍！ 昨夜，AI圈又迎来一次深夜"突袭"。 DeepSeek（深度求索），在未开发布会的情况下，悄然上线了其全新的V3.1版本模型。 架构创新信号：线上模型悄然去除了"R1"（代表深度思考）的标识，并新增了search和think等特殊Token，引发了行业对DeepSeek未来可能采用"混合架 构"的广泛猜测。 <|search_begin|>(id:128796)<|search_end|>(id:128797)(id:128798)(id:129899) 公开的评测数据是"过去时"，而投资决策永远面向"未来时"。当社区和媒体还在为V3.1的性能跑分欢呼时，真正敏锐的资本已经开始对AI赛道的底层逻 ...

Core Viewpoint - OpenAI's founder Sam Altman believes that the U.S. is underestimating the threat posed by China's next-generation artificial intelligence, and that chip regulations alone are not an effective solution [1][3] Group 1: AI Competition - Altman stated that China can develop faster in reasoning capabilities and has strengths in research and product development [3] - The AI competition between the U.S. and China is deeply intertwined, going beyond simple rankings [3] Group 2: OpenAI's Strategic Shift - OpenAI recently released its first open-weight models, gpt-oss-120b and gpt-oss-20b, marking a significant strategic shift from its long-standing closed-source approach [3] - The decision to release open-weight models was influenced by competition from Chinese models, such as DeepSeek and Kimi K2 [3] - Altman emphasized that if OpenAI did not act, Chinese open-source models would gain widespread adoption, making this a significant factor in their decision [3]

Group 1: Photovoltaic Industry - The Ministry of Industry and Information Technology (MIIT) is actively coordinating between power generation companies and local industries to enhance price transmission from manufacturing to power stations, emphasizing a market-oriented and legal approach to eliminate outdated production capacity [1] - The average bidding price for components from China Resources and China Huadian has increased by 5-8% month-on-month, while silicon material companies have proactively limited production, leading to a 10% decrease in silicon wafer inventory over two weeks [1] - The investment sequence indicates a tight supply of silicon materials in Q3, a premium for BC battery technology in Q4, and a simultaneous increase in both volume and price of auxiliary materials such as glass and adhesive films [1][2] Group 2: Solid-State Battery Technology - A breakthrough in solid-state battery technology has been achieved with the introduction of 5μm vapor-deposited lithium anodes, significantly reducing dendrite risk and achieving over 500 cycles with a capacity retention rate above 90% [3] - The cost of 5μm vapor-deposited lithium is projected to drop to 2 million yuan per GWh, compared to 4 million yuan for 20μm rolled lithium foil, indicating a substantial cost reduction in the industry [3] - The solid-state battery market could reach 50-100 billion yuan by 2030, driven by the demand for 100GWh of global solid-state battery production [3] Group 3: Robotics Industry - The T-Link Gen3 technology is reshaping the supply chain with a focus on lightweight materials, energy efficiency, and sensor integration, leading to a re-tendering of motors, reducers, and lead screws [4] - The use of PEEK materials has reduced costs by 30% compared to imports, and the new harmonic magnetic field motors have achieved a 50% reduction in size while doubling power density [5] - The 3D vision solution from Orbbec has a single machine value of 200 USD, and the company has passed factory audits [6] Group 4: Semiconductor and AI Models - The DeepSeek model has been upgraded to V3.1, expanding the context length from 64K to 128K, which is expected to increase demand for GPU memory and HBM [7] - The need for larger training clusters is anticipated to rise by 30%, benefiting semiconductor and storage manufacturers such as Cambricon, Haiguang, and Lanke [7] Group 5: Pharmaceutical Industry - Rongchang Biotech has licensed its ophthalmic drug RC28-E to Japan's Santen Pharmaceutical, marking a shift in domestic innovative drug licensing from popular fields like oncology to specialized areas with differentiated advantages [8] - This collaboration model provides a clear path for value realization in less popular biotech sectors through upfront payments, milestones, and sales sharing, enhancing cash flow and leveraging established commercialization channels [8] Group 6: High-Speed Rail Industry - The China National Railway Group has initiated its second batch of high-speed train tenders for the year, with 210 sets, marking a recent high and exceeding market expectations [9] - This move reinforces the trend of sustained railway investment recovery, with new construction and maintenance peaks positively impacting the performance certainty of core companies in the industry [9]

来源：环球网 【环球网科技综合报道】8月20日消息，DeepSeek日前在Hugging Face上开源了新模型 V3.1-Base。 此外，日前DeepSeek 还发布通知称，线上模型版本已升级至 V3.1，上下文长度拓展至 128k，可通过官 方网页、App、小程序测试，API 接口调用方式保持不变。 就在8月14日，DeepSeek App发布了1.3.0版本，此次更新在修复已知问题、优化文本操作体验的基础 上，首次引入"对话内容生成分享图"功能，为用户提供更便捷、个性化的内容传播方式。（思瀚） ...

就在昨晚，DeepSeek官方悄然上线了全新的V3.1版本，上下文长度拓展到128k。 对于这波更新，大家的热情可谓是相当高涨。 即便还未公布模型卡，DeepSeek V3.1就已经在Hugging Face的趋势榜上排到了第四。 本次开源的V3.1模型拥有685B参数，支持多种精度格式，从BF16到FP8。 综合公开信息和国内大咖karminski3的实测，V3.1此次更新亮点有： 编程能力：表现突出，根据社区使用Aider测试数据，V3.1在开源模型中霸榜。 性能突破：V3.1在Aider编程基准测试中取得71.6%高分，超越Claude Opus 4，同时推理和响应速度更快。 原生搜索：新增了原生「search token」的支持，这意味着搜索的支持更好。 架构创新：线上模型去除「R1」标识，分析称DeepSeek未来有望采用「混合架构」。 成本优势：每次完整编程任务仅需1.01美元，成本仅为专有系统的六十分之一。 值得一提的是，官方群中强调拓展至128K上下文，此前V3版本就已经支持。 | Model | #Total | #Activated | Context | Download | | --- ...

没等到Deepseek R2，DeepSeek悄悄更新了V 3.1。 官方群放出的消息就提了一点，上下文长度拓展至128K。128K也是GPT-4o这一代模型的处理Token的长度。因此一开始，鲸哥以为从V3升级到V 3.1，以 为是不大的升级，鲸哥体验下来还有惊喜。 代码能力与前端审美提升 从开源社区Huggingface上传的模型版本看，模型尺寸达685B，支持 BF16、F8_E4M3、F32 等张量类型，平衡模型的计算精度和效率。 最惊喜的是代码能力提升明显，前端审美也有大幅度提升。我们先看V3.1在代码测试中的变现。 请设计并开发一款结合日历和待办事项（To-Do）的产品，其核心功能应包括： 任务分类与颜色标记：用户能够创建不同类别的任务，并为每个类别分配独特的颜色。当任务被归类后，其在日历视图上应以相应的颜色进行标记，以便 快速识别。短期任务管理：*完成标记： 对于计划在特定日期完成的任务，用户应能将其标记为"已完成"。已完成的任务应在界面上以视觉方式（例如， 划掉、变灰或显示完成图标）清晰区分。*逾期处理： 如果任务未在计划日期完成，系统应提供明确的视觉提示（例如，颜色变化、闪烁或标记为逾 期） ...

上方为DeepSeek-V3-0324开源网页，下方为DeepSeek-V3.1-Base开源网页 开源地址： https://huggingface.co/deepseek-ai/DeepSeek-V3.1-Base 智东西第一时间在网页端对新模型的能力进行了体验，从初步体验结果来看，这一模型在编程（尤其是前端能力）、物理定律理解、创意写 作、数学、回答语气等方面都出现不同程度的提升和变化。 以下是智东西体验的部分案例： 此外，DeepSeek还将App、网页端的"深度思考（R1）"字样改为了"深度思考"，有网友猜测这是融合推理模型与非推理模型的征兆，但 DeepSeek官方尚未发布任何关于这一改动的消息。 左侧为旧版页面，右侧为新版页面 这一模型现已上传至Hugging Face，不过目前仅开源了未经指令微调的Base版本（基础模型），其配置文件、脚本代码和模型权重均可供下 载。与DeepSeek-V3-0324相比，模型参数量、张量类型没有明显变化。 智东西8月20日报道，昨日晚间，DeepSeek在官方群宣布：DeepSeek线上版本模型已升级至DeepSeek V3.1，上下文窗口从原有的64k扩展 ...