AI News AI资讯 8h ago Updated 2h ago 更新于 2小时前 46

8:1 Ke | Zong Fuli's Half-Brother Appointed Chairman of New Company; DeepSeek Rises to No.1 in Global API Calls; iPhone May See Price Hikes Over 1000 Yuan 8点1氪丨宗馥莉同父异母兄弟出任新公司董事长;DeepSeek升至全球调用量第一;iPhone被曝最高或涨价超千元

DeepSeek-V4-Flash surged to #1 in global AI model call volume, with 7.22 trillion tokens weekly, as Chinese models collectively dominated the top 5 rankings for the 14th consecutive week OpenAI's next-gen model Astra internally solved 10 advanced mathematics and theoretical computer science problems at an estimated $2,000 compute cost, signaling rapid progress in AI-driven mathematical reasoning Alibaba released Qwen3.8, a 2.4-trillion-parameter base model with significant improvements in coding DeepSeek-V4-Flash登顶全球AI大模型调用量榜首,中国大模型周调用量连续十四周超美国,占全球总量近半 OpenAI下一代模型Astra以约2000美元算力成本破解十项高等数学与理论计算机科学难题,展现强推理能力 阿里巴巴发布Qwen3.8基座大模型,总参数量达2.4万亿,编程与办公场景能力显著提升,Arena榜单仅次于Claude MiniMax H3正式开源,支持文本/图像/视频/音频全模态理解,可生成最高2K分辨率、15秒带原生立体声音频的视频 欧盟《人工智能法案》透明度规则正式生效,要求AI交互系统明确告知用户身份,AI生成内容须附带机器可读数字标记

68
Hot 热度
65
Quality 质量
62
Impact 影响力

Analysis 深度分析

TL;DR

  • DeepSeek-V4-Flash surged to #1 in global AI model call volume, with 7.22 trillion tokens weekly, as Chinese models collectively dominated the top 5 rankings for the 14th consecutive week
  • OpenAI's next-gen model Astra internally solved 10 advanced mathematics and theoretical computer science problems at an estimated $2,000 compute cost, signaling rapid progress in AI-driven mathematical reasoning
  • Alibaba released Qwen3.8, a 2.4-trillion-parameter base model with significant improvements in coding and professional office workflows, ranking just behind Anthropic's Claude series on the Arena leaderboard
  • MiniMax open-sourced H3, a universal multimodal video generation model capable of producing 2K-resolution, 15-second videos with native stereo audio from text, image, video, and audio contexts
  • The EU AI Act's core transparency rules officially took effect, requiring chatbots to disclose AI interaction and mandating machine-readable digital watermarks on AI-generated media content

Why It Matters

This digest captures a pivotal moment where Chinese AI models are consolidating global market dominance in usage volume while Western models like OpenAI push the frontier on complex reasoning tasks. For AI practitioners, the DeepSeek and Qwen releases demonstrate that open and accessible models are closing the capability gap, while the EU AI Act enforcement marks the beginning of real-world regulatory compliance requirements that all AI developers and deployers must address.

Technical Details

  • DeepSeek-V4-Flash achieved 7.22 trillion weekly tokens (up 13% MoM), leading a top-5 lineup entirely composed of Chinese models; the model's API also launched on China's National Supercomputing Internet for public access
  • OpenAI Astra internally solved 10 breakthrough problems including high-dimensional sphere packing density bounds, Conway's rigid conjecture refutation, arithmetic circuit complexity lower bounds, and the quantum parallel repetition theorem, at approximately $2,000 in token costs via Sol API pricing
  • Alibaba Qwen3.8 features a 2.4-trillion-parameter architecture with notable gains in Coding and Cowork benchmarks, ranking second only to Anthropic's Claude series on the Arena leaderboard; Qwen3.8-Max and Qwen3.8-27B are scheduled for open-source release
  • MiniMax H3 is a fully open-sourced universal multimodal generation system that unifies understanding of text, image, video, and audio contexts, generating videos up to 2K resolution, 15 seconds long, with native stereo audio
  • EU AI Act transparency rules require interactive AI systems (e.g., chatbots) to explicitly inform users they are interacting with AI, and mandate clear labeling plus machine-readable digital watermarks for AI-generated or modified images, video, audio, and text

Industry Insight

  • The sustained dominance of Chinese models in global call volume (28.13 trillion tokens vs. 4.38 trillion for US models) signals a structural shift in AI adoption patterns, suggesting that cost-effective open models are driving enterprise and developer migration away from premium Western offerings
  • OpenAI's Astra results demonstrate that AI is reaching human-level research capability in specialized mathematics, which will accelerate R&D workflows across scientific computing and theoretical CS; practitioners should monitor how these reasoning breakthroughs translate into production model capabilities
  • The EU AI Act enforcement creates immediate compliance obligations for any AI product serving European users—developers must implement disclosure mechanisms and watermarking infrastructure now, as regulatory penalties will follow for non-compliance

TL;DR

  • DeepSeek-V4-Flash登顶全球AI大模型调用量榜首,中国大模型周调用量连续十四周超美国,占全球总量近半
  • OpenAI下一代模型Astra以约2000美元算力成本破解十项高等数学与理论计算机科学难题,展现强推理能力
  • 阿里巴巴发布Qwen3.8基座大模型,总参数量达2.4万亿,编程与办公场景能力显著提升,Arena榜单仅次于Claude
  • MiniMax H3正式开源,支持文本/图像/视频/音频全模态理解,可生成最高2K分辨率、15秒带原生立体声音频的视频
  • 欧盟《人工智能法案》透明度规则正式生效,要求AI交互系统明确告知用户身份,AI生成内容须附带机器可读数字标记

为什么值得看

本文集中呈现了2026年8月初全球AI领域的关键动态,既涵盖中国大模型在全球调用量上的持续领先,也记录了OpenAI、阿里、MiniMax等头部机构在模型能力与开源策略上的最新进展,同时涉及AI监管的重要里程碑,对把握行业技术趋势与政策走向具有重要参考价值。

技术解析

  • DeepSeek-V4-Flash调用量登顶:根据OpenRouter数据,上周全球AI大模型总调用量56.8万亿Token,其中中国模型周调用量28.13万亿Token,连续十四周超过美国。DeepSeek-V4-Flash以7.22万亿Token周调用量位居第一,环比增长13%;前五名均为中国模型,小米MiMo-V2.5、腾讯Hy3、DeepSeek-V4-Pro、智谱GLM-5.2分列二至五位。
  • OpenAI Astra数学突破:Astra内部版本在数学与理论计算机科学领域实现十项突破,涵盖高维球体堆积密度上限、康涅斯刚性猜想推翻、量子平行重复定理等前沿问题,按Sol API费率计算总成本约2000美元,展示了大模型在复杂推理任务上的成本效率。
  • 阿里巴巴Qwen3.8发布:总参数量2.4万亿,在编程(Coding)和专业办公(Cowork)场景能力大幅提升,已上线千问AI平台并接入Agent产品"千问办公";Qwen3.8-Max与Qwen3.8-27B预计开源,Arena榜单中仅次于Anthropic Claude系列。
  • MiniMax H3开源:新一代通用视频模型,支持文本、图像、视频、音频全模态上下文理解,可生成最高2K分辨率、最长15秒、带原生立体声音频的视频,标志着开源视频生成模型在质量与多模态融合上的进一步突破。
  • 欧盟AI法案透明度规则生效:全球首部全面监管AI的法律进入实质执行阶段,要求聊天机器人等交互式系统明确告知用户正在与AI互动,AI生成或修改的多媒体内容须进行清晰标识并附带机器可读数字标记,为AI内容溯源与合规提供法律框架。

行业启示

  • 中国大模型全球影响力持续扩大:中国模型在调用量上连续十四周领先美国,且前五名全部由中国模型包揽,表明国产大模型在开发者生态、应用落地和成本控制方面已形成显著优势,建议持续关注DeepSeek、阿里、腾讯等头部模型的开源策略与商业化路径。
  • AI能力竞争从参数规模转向推理效率:OpenAI Astra以2000美元成本破解十项数学难题,反映出行业竞争焦点正从单纯堆砌参数转向提升模型在复杂推理任务上的效率与性价比,企业应重视模型在专业场景的实用价值而非仅关注基准测试排名。
  • AI监管进入实质性执行阶段:欧盟AI法案透明度规则生效标志着全球AI治理从立法走向落地,科技公司在产品设计和内容生成环节需提前布局合规机制,包括AI身份标识、生成内容水印等技术方案,以应对日益严格的监管要求。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

OpenAI OpenAI LLM 大模型 Product Launch 产品发布 Research 科学研究