DeepSeek 最新版 V3.1 被多名开发者实测发现，会在完全不该出现的地方插入「极 / 極 / extreme」等 token。 | | VolcEngine DeepSeek Reasoner (V3.1) 深度求索 08/25 17:37 | | --- | --- | | ે | 已深度思考(用时5.9秒) | | | | | | var data = [ | | 2 | [0,0,0.5],[0,1,0.5],[0,2,0.5],[0极,3,0.5],[0,4,0.5],[0,5,0.5],[0,6,0.38],[0,7,0.44],[0,8极 | | 3 | [1,0,0.5],[1,1,0极.5],[1,2,0.5],[1,3,0.5],[1,4,0.5],[1,5,0.25],[1,6,0.44],[1,7,0.57],[1,8, | | 4 | [2,0,0.5],[2,1,0.5],[2,2,0.5],[2,3,1.0],[2,4,0.71],[2,5,0.2],[2,6,0.71],[2,7,0.5],[2,8,0.2 | | 5 | [3,0,0.5],[3,1,0.5],[3, ...

本文来自微信公众号：APPSO （ID：appsolution），作者：APPSO，题图来自：AI生成 DeepSeek最新版V3.1被多名开发者实测发现，会在完全不该出现的地方插入"极/極/extreme"等token。 | | VolcEngine DeepSeek Reasoner (V3.1) 深度求索 08/25 17:37 | | --- | --- | | ਨੂੰ | 已深度思考(用时 5.9 秒) | | | | | | var datal = [ | | 2 | [0,0,0.5],[0,1,0.5],[0,2,0.5],[0极,3,0.5],[0,4,0.5],[0,5,0.5],[0,6,0.38],[0,7,0.44],[0,8极 | | 3 | [1,0,0.5],[1,1,0极.5],[1,2,0.5],[1,3,0.5],[1,4,0.5],[1,5,0.25],[1,6,0.44],[1,7,0.57],[1,8, | | 4 | [2,0,0.5],[2,1,0.5],[2,2,0.5],[2,3,1.0],[2,4,0.71],[2,5,0.2],[2,6,0.71],[ ...

Core Viewpoint - The recent rise in chip and AI computing indices is driven by the increasing demand for AI capabilities and the acceleration of domestic chip alternatives, highlighted by DeepSeek's release of DeepSeek-V3.1, which utilizes the UE8M0 FP8 scale parameter precision [2][5]. Group 1: Industry Trends - The chip index (884160.WI) has increased by 19.5% over the past month, while the AI computing index (8841678.WI) has risen by 22.47% [2]. - The introduction of FP8 technology is creating a significant trend in low-precision computing, which is essential for meeting the industry's urgent need for efficient and low-power calculations [2][5]. - Major companies like Meta, Microsoft, Google, and Alibaba have established the Open Compute Project (OCP) to promote the MX specification, which packages FP8 for large-scale deployment [6]. Group 2: Technical Developments - FP8, an 8-bit floating-point format, is gaining traction as it offers advantages in memory usage and computational efficiency compared to previous formats like FP32 and FP16 [5][8]. - The transition to low-precision computing is expected to enhance training efficiency and reduce hardware demands, particularly in AI model inference scenarios [10][13]. - DeepSeek's successful implementation of FP8 in model training is anticipated to lead to broader adoption of this technology across the industry [14]. Group 3: Market Dynamics - By Q2 2025, the market share of domestic chips is projected to rise to 38.7%, reflecting a shift towards local alternatives in the AI chip sector [9]. - The Chinese AI accelerator card market share is expected to increase from less than 15% in 2023 to over 40% by mid-2025, indicating a significant move towards self-sufficiency in the domestic chip industry [14]. - The industry is witnessing a positive cycle of financing, research and development, and practical application, establishing a sustainable path independent of overseas ecosystems [14].

国产算力产业链正稳步走出一条独立于海外生态的可持续路径。 8月26日，芯片指数（884160.WI）探底回升，午盘涨0.02%，近一个月涨19.5%；AI算力指数（8841678.WI）热度延续，午盘涨1.45%，近一个月涨22.47%。 虽然DeepSeek V3.1预告将匹配UE8M0 FP8 Scale参数精度，并引爆FP8及低精度方面热度，但在行业内，该参数已非新事物。 消息面上，DeepSeek上周发布DeepSeek-V3.1，称此次升级是迈向Agent（智能体）时代的第一步。DeepSeek称，DeepSeek-V3.1使用了UE8M0 FP8 Scale参数 精度，并表示UE8M0 FP8是针对即将发布的下一代国产芯片而设计。 芯片指数与AI算力指数近期持续走高背后，是AI浪潮与大模型算力需求剧增下，国产替代加速与供应链多元化路径日渐成熟的趋势。而DeepSeek掷出FP8这 颗 "魔力骰子"，不仅精准切中行业对高效低功耗计算的迫切需求，更直接引发了一场围绕低精度计算的现象级热潮，为国产算力赛道再添一把火。 爆火前的三年成长期 近期，借DeepSeek"东风"，二级市场多家芯片公司与券商机构密 ...

其中，全新BMW X3长轴距版曜夜套装给外观换上新"皮肤"，售价不变。全系车型还增加新车身 漆"个性化定制磨砂纯灰"，冷冽色泽散发锋芒。全新BMW X3长轴距版设计语言着眼于未来家族，原石 切割般的车身带来无可比拟的气势。其轴距达2,975毫米，媲美BMW X5标准轴距。车身每一曲线和棱 角，都经过严格的风洞测试，风阻系数较上代降低7%，驾驶更高效。 颜值换新之外，BMW X家族还将更智能。未来几周内，BMW X1、X3长轴距版搭载的BMW智能 个人助理将接入DeepSeek功能，扩展车机能力边界。同时，第9代BMW操作系统车机生态不断解锁"新 技能"，即将带来常用、好用的新应用和新功能，让数字化体验始终在线。驾驶过程中，车道级导航覆 盖城市主干道；3D视图的车载地图直观呈现精准路况；一线城市更可实现精确到车位的地下停车场导 航。(中国经济网记者 郭跃) "X"是BMW体系中最具进取精神的代表。此次BMW X家族主力成员全面引入"曜夜套装"，为BMW X1、X3长轴距版、X5车身覆上亮黑高光，遍布车身的高亮修饰恰到好处，个性化、运动风双buff叠 满，更具张力的视觉表达呼应客户积极进取、追求豪华品质与时尚格调 ...

据硅基流动消息，硅基流动大模型服务平台已上线深度求索团队最新开源的DeepSeek-V3.1，支持160K 超长上下文。 （文章来源：第一财经） ...

据硅基流动消息，8月25日，硅基流动大模型服务平台上线深度求索团队最新开源的DeepSeek-V3.1。 DeepSeek-V3.1总参数共671B，激活参数37B，采用混合推理架构（同时支持思考模式与非思考模 式）。此外，DeepSeek-V3.1率先支持160K超长上下文，让开发者高效处理长文档、多轮对话、编码及 智能体等复杂场景。 ...

U18M 零 IP8 格式如何在节省算力和内存的情况下提升效率？ U18M 零 IP8 格式通过将权重数据从 128 乘 128 量化块拆分成 128 乘 4 的小 块，从而减少显存占用和计算开销，同时保持计算精度。传统的 IP8 权重需要 大量显存在 128 乘 128 块中反复使用，而新的 U18M 零 IP8 则通过更小的数 据块减少了这些需求。此外，新方法还优化了反向量化过程，进一步节省存储、 显存和计算资源。这些改进使得新格式能够在保持高精度的同时，大幅提高训 练和推理效率。 大厂怎么看 DeepSeek-V3.120250824 摘要 Deepseek 定义 U18M 零 IP8 格式，旨在为国产芯片制定新标准，降低 训练侧显存占用 20%-30%，提升训练效率 30%-40%，并指导下一代 国产芯片设计，有望通过 OCP 扩展为国产芯片的 RP8 协议标准。 U18M 零 IP8 通过拆分量化块减少显存占用和计算开销，优化反向量化 过程，在保持高精度的前提下提高训练和推理效率，并采用混合精度策 略平衡性能与精度，敏感参数保留高精度计算（如 FP16）。 SP8 数据格式将提升国产大模型训练效率， ...

Core Insights - The launch of DeepSeek-V3.1 marks a significant step towards the era of AI agents, with developers now able to build their own intelligent agents [1] - Alibaba's introduction of the Qoder programming platform highlights the competitive landscape in AI programming, with major players like ByteDance and Tencent also entering the market [2] - The AI programming sector is rapidly growing, with at least seven unicorns valued over $1 billion and total funding exceeding 240 billion RMB [2][3] Group 1: Product Developments - DeepSeek-V3.1 achieved a score of 76.3% in Aider coding tests, outperforming competitors like Claude 4 Opus and Gemini 2.5 Pro [1] - Qoder integrates top programming models and can search through 100,000 code files at once, significantly enhancing software development efficiency [1] - Anysphere's Cursor has gained approximately 30,000 enterprise clients and reached an annual recurring revenue (ARR) of over $500 million, showcasing its rapid growth in the AI programming space [3] Group 2: Market Dynamics - The AI programming race has intensified, with major tech companies vying for control over the ecosystem rather than just competing on product features [2] - The potential market for personalized software development could reach up to $15 billion by 2030, driven by reduced costs and barriers to entry in software development [6] - The rise of open-source strategies among domestic companies, such as Qwen3-Coder and DeepSeek-V3.1, is attracting global developers and fostering ecosystem growth [5][6] Group 3: Competitive Landscape - The AI programming sector is characterized by a unique advantage for domestic tech firms, which includes performance catch-up and ecosystem collaboration [4] - The market share of domestic models like Tongyi Qianwen has increased from 5% to 22% in the AI programming field within a month [6] - The competition is not only about faster coding but also about establishing a stronghold in the next wave of AI and computational power [5]

近日，DeepSeek 正式推出 V3.1 版本，完成了一次以"工程实用主义"为核心的全面升级。作为AI算力与 智算解决方案的提供者，英博数科持续关注此次迭代对工具调用、思维链条与系统集成的优化，在不牺 牲原有性能的前提下，实现更稳健、高效、低成本的落地表现。 在经历数轮大规模预训练与强化优化后，DeepSeek 于本次迭代推出V3.1，定位非常明确：在不牺牲主 流任务质量的前提下，把工具调用、思维组织与系统集成做得更稳、更快、更"省"。 概览：一次"以用为先"的增量跃迁 与以往强调纯粹大模型能力不同，DeepSeek V3.1 更像一次"工程化特性"驱动的版本： ·思维模式支持更完整：tokenizer 增加了 4 个与推理/检索相关的特殊 token，配合后训练的策略约束， 使"思考—检索—工具—回答"的链条更可控。 ·工具与代理能力更稳：在函数调用、检索增强、智能代理等场景中，调用意图更明确、参数更规整、 失败重试更克制。 ·"Think" 变体效率提升：DeepSeek-V3.1-Think 的整体回答质量大体对齐DeepSeek-R1-0528，但响应更 快，吞吐与时延表现更友好。 ·更贴近硬件的训 ...