AI NEWSLatest AI NewsArticleClaude Fails to Solve a Century-Old Math Puzzle but Accidentally Rewrites the Fundamental Boundaries of Human Mathematical History!Published in Latest AI NewsTime :Aug 11, 2026Read :3minuteRecently, the field of artificial intelligence has witnessed a remarkable new breakthrough. Claude, an unreleased research version developed by Anthropic, attempted to challenge one of the highest peaks in the mathematical world, the Riemann Hypothesis. Although it did not fully solve this century-old problem, it unexpectedly increased the lower bound of the proportion of zeros on the critical line of the Riemann zeta function from 41.6%, which had been maintained for decades, to 67.2% during the exploration process.This achievement quickly sparked heated discussions in the academic community. It should be clarified that this does not mean the Riemann Hypothesis has been proven to 67.2%. The study does not provide definitive guidance for the hypothesis itself. However, its real significance lies in the new form of scientific research demonstrated by AI. During this multi-round effort, Claude did not rely on traditional search for known answers, but instead orchestrated approximately 60 sub-Agents to form a “virtual research team,” experiencing hundreds of failures, independently reviewing massive literature, and recombining classical methods. Ultimately, it explored innovative boundaries that had never been established before in the open and unknown field of mathematics.Industry experts point out that this breakthrough not only showcases the hard-core capabilities of artificial intelligence in complex, long-range, multi-agent collaboration, and autonomous error verification, but also signals a potential revolution in future research workflows. AI may no longer be limited to solving known problems, but instead, by undertaking intensive trial-and-error processes and cross-referencing vast amounts of literature, it will significantly reduce the marginal search costs for humans in unknown fields, ushering in a new era of scientific exploration.

AI NEWSLatest AI NewsArticleDoubao Work Task Function Update: Control Your Computer Remotely from Your Phone, Breaking Device Barriers for Cross-Device WorkPublished in Latest AI NewsTime :Aug 17, 2026Read :4minuteRecently, ByteDance’s smart assistant Doubao announced a major update to its “Work Tasks” mode - the official launch of the mobile phone remote control computer function. Users just need to complete a simple authorization on the Doubao app on their mobile phone, and they can take over the computer logged in with the same account at any time, handle desktop tasks, breaking the device restrictions in the office scenario, and providing a more flexible solution for mobile work.According to the official introduction, this feature supports retrieving file context in the local computer environment, efficiently handling documents, images, code, tables, and other types of materials, and completing complex computer operations. It has been verified for practicality in multiple high-frequency office scenarios: for example, if a user leaves home and finds that a file on the computer has not been sent, they can remotely search for local files on their phone, complete format conversions such as Pages to PDF, and send them back to the phone; when temporarily receiving modification suggestions but the computer is not nearby, they can remotely instruct the computer to generate summaries and charts and synchronize them to the phone; it also supports linking multiple computers with a single phone, enabling cross-device collaboration, and sending files from different computers to the phone in a unified way.The operation threshold has also been significantly reduced: users just need to update Doubao to the latest version, log in with the same account as the computer, select the “Work Tasks” mode, click “Cloud Computer” to choose the operating environment, receive an SMS verification code to complete the authorization binding, and then quickly connect to the device and start using it.Industry analysts believe that this feature accurately addresses the pain point of professionals who need to urgently handle work without having their computer nearby, further implementing the cross-terminal collaboration capabilities of smart assistants, and is expected to become the new standard in mobile office scenarios.

AI NEWSLatest AI NewsArticleAmazon is being overwhelmed by AI-generated books, and the living space for human authors is being severely squeezedPublished in Latest AI NewsTime :Aug 17, 2026Read :4minuteA new study of over 14,000 self-published e-books on Amazon shows that AI-generated content is quietly reshaping the entire publishing market by leveraging its sheer volume. Data indicates that the size of Amazon’s book catalog increased by 38.3 times from the beginning of 2023 to the first quarter of 2026, while the platform’s quarterly total revenue only grew by 8.9 times during the same period. This means that more and more books are competing for a slowly growing, limited revenue pool, leading to a noticeable “dilution effect” in the market.What is concerning is that this impact is not limited to low-quality content; even ordinary books without any detectable AI text have also been affected. In the eight major book genres covered in the statistics, the income per book for non-AI texts declined in seven categories. The study points out that books with a large amount of AI content account for 20% of the catalog but contribute only 12.1% and 11.3% to sales and revenue, respectively. Conversely, books without detectable AI text still account for 72.5% of mainstream revenue. However, as AI books continue to climb to the top of bestseller lists, the proportion of AI works in the new Top 25 list has reached as high as 31%.Additionally, the analysis found that some prolific authors have significantly increased their output using AI, with some authors earning millions of dollars from just a few books. In terms of language features, best-selling AI books show a high degree of overlap with existing published works, relying much more on rare expressions than typical literary works.Industry experts believe that this study provides key empirical data for ongoing AI copyright lawsuits. In the past, AI companies often defended themselves by citing human reading and learning as justification. However, evidence of market dilution and linguistic overlap may have a profound impact on future judicial judgments regarding fair use.

AI NEWSLatest AI NewsArticleAlibaba Launches AI Music Model HappyShrimp 1.0, Natural Language Creation Breaks Music Production BarriersPublished in Latest AI NewsTime :Aug 17, 2026Read :4minuteOn August 17, Alibaba officially launched the first version of its self-developed AI music generation model, HappyShrimp (Chinese name: Kuaile Xiaomi). This model breaks through the professional barriers of traditional AI music tools that rely on “tag-based input,” and for the first time realizes a full-process creation driven by natural language in daily life, from lyric writing, composition, arrangement to singing, allowing ordinary users to produce original music with high industrial completion.Different from AI music tools on the market that require stacking professional terms such as “Lo-fi R&B” or “808 drum patterns,” HappyShrimp regards music as a “special language with grammar, semantics, and context.” It uses end-to-end whole-song generation technology, avoiding problems like lyrical and musical misalignment and fragmented listening experiences caused by multi-module combinations. In test cases, when users input “monologue of a ghost being abandoned,” the model autonomously completed the entire structural design, from organ atmosphere setting, male voice slow storytelling, to emotional climax, with lyrics and melodies perfectly matching and an immersive experience.Its core capability also lies in precise control over complex instructions: when given a composite request such as “70s soul music, Black male voice, warm analog recording texture,” the generated segment presented a thick and realistic vocal at 18 seconds, with smooth vibrato close to that of a professional singer; for scenarios such as “car movie theme song, male and female rap duet,” the work not only matched Hans Zimmer’s film score style but also recreated the tension of a race track through string progression and intertwined male and female voices.On its first day, HappyShrimp announced a strategic cooperation with Tahe Music Group, and the two parties will explore areas such as building the music industry ecosystem, operating AI music platforms, and co-creating with musicians. In addition, the model will be showcased at the 2026 Aanaia Xiami Music Festival from August 28 to 30, and will release thematic ambient music inspired by the festival on-site.Currently, HappyShrimp is available both domestically (https://www.happyshrimp.cn/) and internationally (https://www.happyshrimp.ai/). New users can claim large amounts of free credits for trial use. Industry insiders believe that the launch of this product will further expand the creative boundaries of AI music and bring new production models to the music industry.

ENLatest AI NewsApple is preparing to launch! It is secretly planning to introduce paid real-time news with major publishersApple is in talks with several news publishers to introduce real-time news and various current information retrieval capabilities for the next generation of Siri AI, and pay for the content. Sources said that the cooperation does not use the industry’s common fixed licensing fee, but tends to an innovative model of paying by usage, to empower the new voice assistant to access real-time information.5 hours ago11.5KAlibaba Launches AI Music Model HappyShrimp 1.0, Natural Language Creation Breaks Music Production BarriersOn August 17, Alibaba launched the AI music generation model HappyShrimp (Happy Shrimp) 1.0. It breaks the traditional tag-based input barriers, allowing users to drive the entire process of lyric writing, composing, arranging, and singing using colloquial natural language, enabling ordinary users to produce high-quality original music. Unlike tools that simply pile up professional terms, it treats music as a creative process with grammar and semantics.6 hours ago12.4KZhou Hongyi sends an internal letter announcing that 360 has launched the AI-native organizational transformation, with annual management performance linked to AI achievementsThe founder of 360, Zhou Hongyi, sent an internal letter announcing the company’s full transition to an AI-native organization, requiring all levels of managers to cultivate “super individuals” and build “super teams.” He stated that this is not just a regular leadership upgrade, but a new entrepreneurship for 360 using AI, marking further enhancement of the “All in Agent” strategy.6 hours ago11.6KAmazon is being overwhelmed by AI-generated books, and the living space for human authors is being severely squeezedA study of over 14,000 Amazon self-published e-books finds AI-generated content is reshaping the publishing market by volume. From early 2023 to Q1 2026, catalog titles surged 38.3x while quarterly revenue grew only 8.9x, causing a dilution effect and concerns for creative ecology…..7 hours ago11.1KDoubao Work Task Function Update: Control Your Computer Remotely from Your Phone, Breaking Device Barriers for Cross-Device WorkThe “Doubao Work Task” mode has been updated with a new feature allowing users to remotely control their computers from their mobile devices. After authorizing through the mobile app, users can take over desktop tasks on their computer account at any time and access local file contexts to efficiently handle documents, images, code, spreadsheets, and more, breaking through device limitations and offering a more flexible solution for mobile work.7 hours ago10.9KU.S. Plaintiff Hid AI Prompt Words in Complaint: Judge Warns Against Prompt Injection Infiltrating CourtsA self-represented plaintiff in Connecticut, USA, concealed hidden instructions in court documents with white text on a white background, only readable by AI, attempting to influence the case handling. It became the first case of “prompt injection” in the U.S. judicial system. However, the state court did not actually use AI to review the materials, so the method did not work, and the party was penalized for abusing the procedure.7 hours ago11.2KAliQwen-Audio-3.0 Series Speech Models Launched on Tongyi AI Platform, Winning a Grand Slam in International EvaluationsAlibaba has fully launched the Qwen-Audio-3.0 series of speech models on the Tongyi AI platform, including three types: speech recognition, speech synthesis, and real-time voice interaction. Officially, the series performed exceptionally well in the July 2024 speech ranking list by the authoritative evaluation platform Artificial Analysis, securing first place in all three categories: speech recognition, real-time interaction, and speech synthesis.7 hours ago11.2KAlipay: Merchant Smart Agents Are Actively Integrating with Abao, Some Development Demands Have Been Scheduled Until Next YearOn August 17, the first smart agent business ecosystem conference in China was held in Hangzhou. Alipay revealed that the super service smart agent “Abao” has completed more than 10,000 services with AI, supporting five mobile phone brands and 16 car manufacturers; it also launched a full-stack smart agent business infrastructure and the AHA protocol, collaborating with more than 20 companies such as Qwen, Huawei, OPPO, BYD, and Geely to build an interconnected ecosystem.7 hours ago11.0KGoogle AI Makes a Big Move: Image and Video Watermarks Can Finally Be Removed!Google introduces a feature letting users turn off visible watermarks when generating images, videos, and music. Revealed by Gemini lead Josh Woodward, it will roll out to models like Nano Banana, Omni, and Lyria, first on Gemini web and Google video products…..8 hours ago12.2KThe AI Industry Faces a Cost-Cutting Trend: Companies Are Becoming More Sensitive to the Costs of Large Models, and They Clearly Have No Expectations for Mainstream

MiniMax Music 3.0 – MiniMax 开源的音乐生成模型AI工具3天前发布AI小集03MiniMax Music 3.0是什么MiniMax Music 3.0 是 MiniMax推出的下一代开放权重音乐生成模型。模型采用 8B Global LLM与 0.6B Local LLM 的分层架构，结合 Flow Matching + Flow-VAE 连续隐状态合成技术，可根据歌词和结构化音乐描述生成最长 5 分钟、32kHz 立体声的完整歌曲。模型支持精细的段落控制，能在长音频中保持主题、节奏与歌声身份的一致性，实现从创作意图到高保真成品的端到端生成。MiniMax Music 3.0的主要功能完整歌曲生成：支持生成最长 5 分钟的完整歌曲，涵盖前奏、主歌、副歌、桥段、尾奏等标准结构。歌词驱动创作：输入歌词与音乐风格描述，即可生成包含人声演唱、器乐编曲的完整作品。结构化段落控制：支持[Verse]、[Chorus]、[Bridge]等标签，精确控制歌曲各段落的情绪与编曲变化。长程一致性保持：在长达数分钟的音频中稳定维持音乐主题、节奏、歌声身份与编曲推进。高保真音频输出：生成 32kHz、16-bit 立体声 WAV 文件，音质接近专业录音室水准。提示词自动增强：内置music-caption-rewriter工具，可将简短描述扩展为包含全局元数据、人声细节与编曲规划的专业结构化提示。MiniMax Music 3.0的技术原理分层自回归架构（Hybrid-LM）：模型由 8B Global LLM 与 0.6B Local LLM 协同工作。Global LLM 基于 Qwen3 初始化，逐帧预测首个 RVQ 语义码本，负责建模歌曲的长程结构与语义；Local LLM 则在每帧内预测其余声学码本，恢复细粒度音色与音质。两者通过联合训练实现宏观结构与微观声学的高效分离与协作。多层残差向量量化（RVQ）：采用 8 层 RVQ 对音乐信息进行分层表征。首层语义码本含 16,384 个条目，捕捉核心音乐语义与结构；其余 7 层声学码本各含 1,024 个条目，逐级编码残差声学细节。训练时先独立优化语义码本建立稳定骨架，再联合训练全部码本，以平衡语义容量、生成稳定性与细节重建。连续隐状态合成：推理融合 Global 与 Local LLM 的最终连续隐状态，作为条件输入 2.4B Flow Matching 模块映射至 VAE 潜空间，再由 123M Flow-VAE 解码为波形。保留了更丰富的高维声学信息，显著提升人声发音准确度、乐器物理真实感与长时一致性。结构化描述框架：引入精细时粒度的结构化描述体系，将单一全局标签扩展为包含 Global Metadata（流派、BPM、调性）、Vocal Details（音色、唱法、和声）与 Arrangement（乐器演进、段落变化）的三段式描述，使主观创作意图转化为模型可执行的专业编曲指令。微信关注回复“开源”，加入AI开源项目交流群如何使用MiniMax Music 3.0环境准备：确保本地拥有两张支持 CUDA 的 NVIDIA GPU，配置好 Python 环境以及 SGLang-Omni 推理框架的依赖。下载模型：执行hf download MiniMaxAI/MiniMax-Music3 –local-dir /path/to/minimax_ttm命令，从 HuggingFace 拉取完整的模型权重到本地目录。启动服务：运行sgl-omni serve –model-path MiniMaxAI/MiniMax-Music3 –port 8000启动推理服务，其中 GPU 0 负责 Qwen3 与八层 RVQ 自回归生成，GPU 1 负责 Flow Matching 与 Flow-VAE 波形解码。准备输入：在 API 请求的input字段中填入带段落标签（如[Verse]、[Chorus]）的歌词，在instructions字段中填入音乐风格、情绪与编曲描述，两者共同作为生成条件。发送请求：向http://127.0.0.1:8000/v1/audio/speech发送 POST 请求，设置model为minimax_ttm、response_format为wav、max_new_tokens为 9000，即可返回生成的 32kHz 立体声 WAV 音频。提示增强（可选）：安装并调用music-caption-rewriterskill，将简短的音乐描述自动扩展为包含全局元数据、人声细节与段落编曲的结构化提示，以提升生成精度与可控性。MiniMax Music 3.0的核心优势完整长曲生成：原生支持最长 5 分钟完整歌曲生成，在长音频中稳定保持音乐主题、节奏、歌声身份与编曲推进的一致性，实现真正的”一首完整歌”。精准理解创作意图：通过结构化描述框架（Structured Caption）将主观创意转化为可执行的专业编曲指令，避免生成内容偏离原始需求，并配备提示增强工具降低专业门槛。高保真音频渲染：采用连续隐状态合成（融合 LLM 隐状态 → Flow Matching → Flow-VAE），而非传统离散 token 解码，输出 32kHz 立体声，乐器物理真实感更强、混音更通透。精细可控的编曲：支持歌词段落标签与三段式结构化描述（全局元数据、人声细节、段落编曲），可对情绪演进、乐器进出、人声技法进行精确时序控制。自然的人声表现：显著减少高频数字伪影，改善发音准确度、呼吸感与和声层次，使生成的人声更接近真实演唱而非合成效果。MiniMax Music 3.0的项目地址项目官网：https://www.minimax.io/blog/minimax-music-3-0-next-generation-open-weights-production-ready-versatile-music-modelGitHub仓库：https://github.com/MiniMax-AI/MiniMax-Music3MiniMax Music 3.0的同类竞品对比对比维度MiniMax Music 3.0Suno v5开发商MiniMaxSuno AI模型性质开放权重（可本地部署）闭源（仅云端使用）技术架构8B Global LLM + 0.6B Local LLM + Flow Matching + Flow-VAETransformer-based（具体参数未公开）最长生成时长5 分钟~5 分钟（v5.5 支持最长 8 分钟）输出音质32kHz / 16-bit 立体声 WAV44.1kHz 立体声（Pro）/ 48kHz 24-bit（Premier）歌词输入支持，带结构化段落标签支持，带元标签（[Verse] 等）音乐描述精度结构化 Caption（流派、BPM、调性、段落编曲三段式精细控制）自然语言风格描述（建议 200 字符以内）部署方式本地双 CUDA GPU（SGLang-Omni）纯云端（官网 / App）MiniMax Music 3.0的应用场景独立音乐人快速 Demo 制作：音乐人输入原创歌词与风格描述，可在数分钟内生成包含人声与编曲的完整 Demo，大幅降低创作前期的试错成本。短视频与播客定制化配乐：内容创作者根据视频情绪曲线生成主题曲或背景音乐，实现从”找版权音乐”到”定制专属音乐”的转变。游戏与互动娱乐动态配乐：开发者将模型接入游戏引擎，根据玩家实时行为动态生成适配场景氛围的长时音乐，增强沉浸感。广告与品牌营销音频内容：广告团队依据品牌调性描述快速产出多版本广告歌与 Jingle，支持 A/B 测试并避免版权纠纷。音乐教育与创作教学辅助：教师用模型演示不同流派、结构与编曲技法的实际效果，学生通过调整结构化描述直观理解音乐制作原理。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇dots3-note preview - 小红书开源的多模态 MoE 模型下一篇OnSolo - 腾讯推出的 AI 原生创作平台相关文章Gemma 2 – 谷歌DeepMind推出的新一代开源人工智能模型AI小集4语鲸 – AI阅读工具，基于深言科技的大模型能力AI小集6GeneBench-Pro – OpenAI 推出的计算生物学研究级基准测试AI小集2GuideGeek – AI旅行助手，输入时间、预算、兴趣点等快速生成详细安排AI小集2GPT Image 1.5 – OpenAI最新推出的图像生成模型AI小集3LipDub – 初创公司Captions推出的海外AI配音工具，实时转换声音口型同步AI小集3暂无评论再想想发表评论暂无评论…热门工具Loomy即梦AISekoAiPPT秘塔AI搜索妙呀堆友Agent美图设计室绘蛙AIupdream办公小浣熊秒哒最新收录OnSoloEpixa知漫剧QwenPawRenoiseLawbot脱敏猫最新文章秒悟Meoo – 1分钟生成前后端完整的网站，真0门槛开发！4小时前HiDream-O1-World – 智象未来推出的全模态交互式世界模型4小时前DeepSeek Harness Desktop – 开源的 DeepSeek Harness 桌面端4小时前dsh-TUI – 开源的 DeepSeek Harness 终端界面插件10小时前DSH Desktop – 开源的 DeepSeek Harness 本地桌面应用12小时前WorkSwarm – openJiuwen 推出的蜂群办公智能体1天前Happy Shrimp – 阿里 ATH 事业群推出的 AI 音乐平台1天前Copula Lab – AI 数据基础设施平台，提供专家级训练数据1天前Qwen3.8-27B – 阿里千问团队开源的新一代 AI 大模型2天前OnSolo – 腾讯推出的 AI 原生创作平台2天前dots3-note preview – 小红书开源的多模态 MoE 模型3天前GLM-5.3 – 智谱最新开源的基座大模型，编程能力提升50%3天前Gemini 3.7 Flash – Google DeepMind 推出最新的主力模型3天前wigolo – 开源的本地搜索工具，支持接入多种编程 Agent3天前Sonilo – AI 视频配乐与音效生成平台，自动分析画面节奏3天前

Gemini Spark – 谷歌推出的个人 AI AgentAI工具3个月前发布AI小集03Gemini Spark是什么Gemini Spark 是 Google 推出的 24/7 个人 AI Agent，基于 Gemini 3.5 模型与Antigravity框架运行，深度集成 Gmail、Docs、Slides 等 Workspace 工具。工具能在设备关闭后仍于云端后台持续执行任务，支持设置周期性触发器、学习新技能并创建完整工作流，将 Gemini 从问答助手升级为可主动代劳的数字代理。Gemini Spark的主要功能全天候后台任务执行：Spark 作为云端 Agent，即使用户关闭笔记本或锁定手机，仍能在后台持续完成复杂任务。周期性任务与触发器：自动解析月度信用卡账单，标记新增或隐藏的订阅费用；按设定条件持续监控指定信息源。可教学的新技能：用户可指令 Spark 检查特定邮件来源（如学校通知），提取关键截止日期，生成摘要发送给指定联系人。完整工作流创建：自动整合分散在邮件与聊天中的会议笔记，生成格式化的 Google Docs 报告，并起草项目启动邮件。MCP 扩展连接：已支持 Canva、OpenTable、Instacart 等第三方服务，未来数周内将通过 MCP 协议直接代用户完成跨平台操作。如何使用Gemini Spark目前 Gemini Spark 处于 Beta 阶段，仅面向美国地区的 Google AI Ultra 订阅用户逐步开放，可在Gemini App设置中查看并启用。Gemini Spark的核心优势真·全天候在线：区别于对话窗口内响应的助手，Gemini Spark 可在用户离线时持续推进任务进度。深度原生集成：与 Google Workspace 无缝打通，数据读写权限与协作能力远超外挂式插件。可教学进化：用户通过自然语言即可训练 Spark 掌握专属工作流，无需编程或复杂配置。多模态任务链：支持从信息收集、内容生成到跨应用执行的端到端闭环，非单点工具调用。渐进式权限控制：用户自主决定 Agent 的开启状态、连接范围与操作边界，隐私与可控性并重。Gemini Spark的同类竞品对比对比维度Gemini SparkClaude Computer UseOpenAIOperator所属公司GoogleAnthropicOpenAI核心定位24/7 云端个人 AI Agent桌面操控型 AI Agent云端浏览器代理运行环境云端后台（设备可离线）本地 macOS（需设备在线）云端沙盒浏览器后台持续执行支持，关闭设备仍运行不支持，随会话结束停止有限，长任务易中断桌面应用操控夏季通过 macOS App 支持已支持完整桌面控制不支持，仅限浏览器内办公套件集成原生深度集成 Workspace连接器 + 屏幕控制 fallback无原生办公集成OSWorld 任务成功率未公开（基于 Gemini 3.5）约 72%–73%约 38%–43%人在回路确认高风险操作前请求确认频繁请求确认频繁请求确认，易冻结Gemini Spark的应用场景个人财务管理：自动追踪订阅服务扣费异动，每月生成账单分析报告并推送提醒。家庭事务协调：持续监控学校邮件，汇总活动与截止日期，每日向父母双方发送结构化摘要。职场项目启动：自动收集会议纪要与邮件讨论，整理为项目文档并起草团队通知邮件。跨平台内容创作：调用 Canva 生成配图，结合 Docs 撰写文案，通过 Gmail 发送完整营销方案。日常行程管理：基于 Calendar 与邮件内容，主动识别冲突并建议调整，甚至代为预订餐厅。# AI工具©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇Gemini 3.5 Flash - Google 推出的新一代 AI 大模型下一篇Gemini Omni Flash - 谷歌推出的多模态视频生成模型相关文章CHANGER – AI换头技术，将演员头部无缝集成到目标身体AI小集3EasyControl Ghibli – 免费生成吉卜力风格图像的 AI 模型AI小集3Voice-Pro – 开源AI音频处理工具，集成转录、翻译、TTS等一站式服务AI小集4Lanta AI – AI内容创作平台，支持视频和图像处理AI小集2滴文 – AI标书制作平台，提供定制化标书内容生成服务AI小集2CleanS2S – 流式语音到语音交互智能体原型，同时进行听和说AI小集2暂无评论再想想发表评论暂无评论…

Gemini 3.7 Flash – Google DeepMind 推出最新的主力模型AI工具3天前更新AI小集02Gemini 3.7 Flash是什么Gemini 3.7 Flash 是 Google DeepMind 推出的新一代主力 AI 模型。模型专为 Coding 与 Agent 工作流设计，在软件工程、网页开发和企业自动化等基准上大幅跃升，部分指标超越Claude Sonnet 5与GPT-5.6 Terra。模型首发价仅为 3.6 Flash 原价的一半，定位高频 Agent 与企业工作流的”性价比杀手”。Gemini 3.7 Flash的主要功能软件工程：支持生成高质量生产级代码，在 FrontierCode 与 DeepSWE 基准上大幅提升，首遍准确率与调试能力显著增强。Web 开发：支持从截图、图片或完整 Design System 输入，一键生成高保真交互式网页，Code Arena Elo 评分领先竞品。Agent 自动化：执行复杂多步骤企业工作流，自动规划、调用工具并在遇到障碍时自适应调整，AutomationBench 表现大幅跃升。知识密集型工作：处理金融、法律、生物科学等领域的复杂文档推理，GDP.pdf 等长文档理解能力显著增强。多模态创作：配合 Nano Banana 将文字提示实时转化为可玩的 3D 游戏，或将静态 PDF 转化为带交互图表的动态网页。Gemini Spark 集成：作为 24/7 个人 Agent 的后台模型，驱动 Google Workspace 内的文件整理、邮件起草与状态更新等自动化操作。Gemini 3.7 Flash的技术原理算法迭代与快速演进机制：模型针对 Coding 和 Agent 场景的后训练优化与能力定向增强，非单纯依赖预训练规模扩张，从而在保持 Flash 系列速度优势的同时，实现软件工程与知识工作质量的跨越式提升。多步骤规划与严谨推理执行：模型内置更主动的推理资源分配机制，面对复杂任务时会投入更多计算资源进行多步骤规划与工具调用，在执行过程中遇到路障时自适应调整策略、澄清意图；这种”更严谨的执行力”大幅减少了人工监督需求与反复重试次数，是其 Agent 与企业自动化能力跃升的核心技术支撑。原生多模态统一理解：3.7 Flash 具备跨文本、音频、图像、代码和视频的统一理解能力，使其能在 3D 游戏生成、机器人训练闭环和文档智能转换等场景中，将视觉输入（如截图、设计稿）与语言指令融合处理，实现从静态 PDF 到交互网页、从文字描述到可运行 3D 资产的端到端生成。Agent 工作流与 Sub-Agent 编排：针对高频 Agent 场景进行底层优化，通过改进的指令遵循精度与路障适应能力，支持主模型调度多个 sub-agents 协同完成任务；编排能力使其在 Web 开发等场景中可并行处理布局生成、交互逻辑编写与视觉组件渲染，成为支撑长期在线、大规模调用的企业自动化后端基础设施。如何使用Gemini 3.7 Flash开发者接入：开发者可通过 Gemini API 或 Google AI Studio 直接调用模型。快速体验：在Google AI Studio中可零成本创建项目并选择 gemini-3.7-flash 模型，快速测试文本、代码与多模态能力。移动集成：Android Studio 已集成该模型，开发者可在移动应用开发流程中获得代码辅助与生成功能。Agent 编排：通过Google Antigravity这一 Agent-first 开发环境，可编排 sub-agents 构建复杂自动化工作流。企业部署：企业用户可通过 Gemini Enterprise Agent Platform 自定义业务流程与工具集成，实现规模化高频调用。企业使用：企业员工也可直接通过 Gemini Enterprise App 统一入口，用内置的 3.7 Flash 能力处理文档与自动化任务。Gemini 3.7 Flash的核心优势极致迭代节奏：Flash 系列从 3.5 进化到 3.7 仅用了 3 个月，3.6 到 3.7 更是压缩到 3 周，Google 对开发者反馈和市场竞争的响应速度在主流模型中无出其右。Coding 能力跃升：FrontierCode 1.1 Main 从 34.4% 提升至 43.6%，DeepSWE v1.1 从 49% 跃升至 65.3%，首遍代码准确率与长周期软件工程能力均获得跨越式增强，已具备生产级主力模型的代码输出质量。Agent 与企业自动化绝对领先：AutomationBench 得分 30.4%，不仅较自身上一代近乎翻倍，更大幅甩开 Claude Sonnet 5（10.7%）和 GPT-5.6 Terra（23.6%），成为当前企业工作流自动化场景的最强模型。部分基准超越旗舰竞品：在 FrontierCode 1.1（43.6%）和 WebDev Code Arena（1588 Elo）两项关键 Coding 基准上，3.7 Flash 已反超 Claude Sonnet 5 和 GPT-5.6 Terra，打破了”Flash 只做轻量辅助”的刻板印象。极致性价比壁垒：推广价输入 $0.75、输出 $3.75 每百万 Tokens，仅为 3.6 Flash 原价的一半，更是 Claude（$2/$10）和 GPT（$2/$12）的三分之一到四分之一；在高频 Agent 和企业级工作流中，这一成本优势随调用量指数级放大。原生多模态端到端创作：支持文字直接生成可玩的 3D 游戏、截图/Design System 一键转化为交互网页、静态 PDF 自动变为带实时图表的动态数据故事，创作链路完整且无需多模型拼接。Gemini 3.7 Flash的项目地址项目官网：https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-3-7-flash/Gemini 3.7 Flash的同类竞品对比对比维度Gemini 3.7 FlashGPT-5.6 Terra发布方Google DeepMindOpenAI定位Coding + Agent 主力模型通用旗舰模型FrontierCode 1.143.6%🥇41.3%DeepSWE v1.165.3%69.6%🥇Code Arena Elo1588🥇1523AutomationBench30.4%🥇23.6%Terminal-bench 2.185.8%87.4%🥇GDP.pdf34.0%🥇24.7%输入价格 / 1M$0.75🥇$2.00输出价格 / 1M$3.75🥇$12.00迭代节奏3 周一更数月周期核心优势性价比 + Agent 自动化绝对 Coding 性能Gemini 3.7 Flash的应用场景智能 Coding 助手：自动生成生产级代码、调试复杂 Bug、完成长周期软件工程任务，首遍准确率较上一代大幅提升。交互式 Web 开发：从截图或 Design System 输入一键生成高保真交互网页，支持多 agent 协作与视差动效。企业工作流自动化：驱动 RPA 与业务系统，自动完成跨平台数据录入、审批流转、报告生成等高频重复任务。3D 游戏与内容生成：配合 Nano Banana 将文字描述实时转化为可玩的 3D 游戏，动态生成角色、道具与纹理。文档智能转换：将静态 PDF 年报、研报自动转化为带实时图表、数据汇总与交互功能的动态网页或数据故事。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇wigolo - 开源的本地搜索工具，支持接入多种编程 Agent下一篇GLM-5.3 - 智谱最新开源的基座大模型，编程能力提升50%相关文章UniTok – 字节联合港大、华中科技推出的统一视觉分词器AI小集2Moises – 音乐人的AI应用，智能分离人声和各种乐器轨道AI小集2Speechelo – AI文本转语音工具，支持多语言转换、调整语速、音调、情感表达AI小集2Flame – 开源的多模态前端代码生成模型AI小集3Epismo – AI工作流管理与协作平台，支持AI技能封装AI小集2Matrix-Game 3.0 – 昆仑万维推出的实时交互式世界模型AI小集3暂无评论再想想发表评论暂无评论…热门工具Loomy即梦AISekoAiPPT秘塔AI搜索妙呀堆友Agent美图设计室绘蛙AIupdream办公小浣熊秒哒最新收录OnSoloEpixa知漫剧QwenPawRenoiseLawbot脱敏猫最新文章秒悟Meoo – 1分钟生成前后端完整的网站，真0门槛开发！4小时前HiDream-O1-World – 智象未来推出的全模态交互式世界模型4小时前DeepSeek Harness Desktop – 开源的 DeepSeek Harness 桌面端4小时前dsh-TUI – 开源的 DeepSeek Harness 终端界面插件10小时前DSH Desktop – 开源的 DeepSeek Harness 本地桌面应用12小时前WorkSwarm – openJiuwen 推出的蜂群办公智能体1天前Happy Shrimp – 阿里 ATH 事业群推出的 AI 音乐平台1天前Copula Lab – AI 数据基础设施平台，提供专家级训练数据1天前Qwen3.8-27B – 阿里千问团队开源的新一代 AI 大模型2天前OnSolo – 腾讯推出的 AI 原生创作平台2天前MiniMax Music 3.0 – MiniMax 开源的音乐生成模型3天前dots3-note preview – 小红书开源的多模态 MoE 模型3天前GLM-5.3 – 智谱最新开源的基座大模型，编程能力提升50%3天前wigolo – 开源的本地搜索工具，支持接入多种编程 Agent3天前Sonilo – AI 视频配乐与音效生成平台，自动分析画面节奏3天前

Grok Bot – SpaceXAI 推出的 AI Agent 系统，多 Bot 并行协作AI工具4天前发布AI小集05Grok Bot是什么Grok Bot 是 SpaceXAI 推出的 AI Agent 系统，能登录用户的账号像真人一样操作网站和工具，24 小时在云端替用户完成实际工作。系统支持多 Bot 并行协作、录制工作流自动执行，甚至 Bot 可创建新 Bot 组建团队。原型来自 Cursor 内部代号「Sand」的 Agent 产品，现架在 Cursor 基础设施上。Grok Bot的主要功能账号代操作：直接登录已有的应用和网站，在界面层像真人一样点击、填写、推进任务，无需 API 或 MCP 支持。云端常驻运行：每个 Bot 拥有专属小型云计算机，24 小时后台运行，不依赖你的设备是否开机，可凌晨自动拉取数据、整理报告。多 Bot 并行协作：可同时创建销售 Bot、招聘 Bot、费用 Bot、Bug 复现 Bot 等，各自独立处理不同任务。Bot 间自主通信：Bot 之间可直接对话、传递任务、协商分工；管理 Bot 可汇总子 Bot 进展，只在关键节点找你确认。Bot 创建 Bot：可指令 Bot 自行”组建团队”，自动创建新的子 Bot 并分配职责。工作流录制与演化：观察你的操作过程并保存为模板，支持在执行中纠正，工作流会随使用不断优化。跨端同步：支持 macOS 桌面端和 iOS，任务进度多端同步。如何使用Grok Bot访问下载：访问Grok Bot官网 https://x.ai/bot，下载 macOS 或 iOS 客户端。登录订阅：用已订阅 Cursor Teams Premium（$120/月）、Cursor Ultra（$200/月）或 SuperGrok Heavy（$300/月）的账号登录。创建 Bot：点击创建新 Bot，为其命名、设置职位头衔和任务描述。分配任务：给 Bot 指定具体工作，如销售外联、招聘筛选、费用管理或 Bug 复现。自动执行：Bot 自动登录指定的应用或网站，像真人一样在界面层操作并自主推进任务。人机协作：Bot 在云端 24 小时运行，仅在需要人类判断的关键节点回来找你确认。多 Bot 协同：可同时创建多个 Bot 并行运行，它们会互相通信、协商分工，甚至自动创建子 Bot 组建团队。Grok Bot的核心优势无接口限制：无需 API 或 MCP，直接登录应用和网站，在界面层像真人一样点击、填写、推进任务。云端常驻：每个 Bot 拥有专属云计算机，24 小时后台独立运行，不依赖你的设备是否在线。多 Bot 协同：可同时部署多个 Bot 处理不同任务，它们之间自主通信、协商分工、协同推进复杂工作流。工作流演化：观察并保存你的操作步骤为模板，支持执行中纠正，工作流随使用持续优化。自组织团队：可指令 Bot 自行组建团队，自动创建子 Bot 并分配职责，实现层级化管理。闭环执行：Grok Bot 能推进到任务最后一步，只在关键节点找你确认。Grok Bot的同类竞品对比对比维度Grok BotOpenClaw运行模式云端常驻，每个 Bot 拥有专属小型云计算机，24 小时后台独立运行，不依赖用户设备在线。本地优先，任务在用户本地设备或自托管服务器上执行，设备关机则任务中断。部署架构中心化托管，Bot 在 xAI 云端运行，用户通过客户端远程查看进度和接收通知。去中心化自托管，数据不经过第三方服务器，支持 Docker 一键部署，隐私可控性更强。交互入口专用桌面客户端（macOS）+ iOS App，Bot 以独立聊天窗口形式存在，支持多 Bot 并行会话。支持 Telegram、Discord、Slack 等 50 多个消息平台接入，通过聊天指令触发任务。模型支持深度绑定 Grok 系列模型，底层由 xAI 自研模型驱动，模型能力统一但不可切换。可自由切换 OpenAI、Anthropic、Google 及本地开源模型，用户可按任务需求灵活配置。工具操作直接登录用户账号，在应用界面层像真人一样点击、填写、操作，无需 API 或 MCP 支持。依赖 MCP（Model Context Protocol）和 API 连接工具，对无接口的传统软件支持有限。多 Agent 协同原生支持多 Bot 并行运行，Bot 之间可直接通信、协商分工、自动创建子 Bot 组建团队。单次会话为主，不同任务间缺乏持久化协同，多 Agent 协作需手动串联。工作流能力支持录制用户操作并保存为模板，Bot 可自主按模板执行并随纠正不断优化，实现工作流演化。通过预设指令和脚本定义工作流，缺乏”观察-学习-自动优化”的闭环演化机制。安全与隐私账号凭证托管在 xAI 云端，Bot 代登录操作，用户需信任平台的数据安全策略。本地运行，账号凭证和任务数据保留在用户侧，不经过第三方服务器，隐私可控性更高。定价门槛需订阅 Cursor Teams Premium（$120/月）、Cursor Ultra（$200/月）或 SuperGrok Heavy（$300/月）。开源免费，仅需承担自托管的服务器成本，对个人开发者和小团队门槛极低。Grok Bot的应用场景销售外联：夜间自动研究客户账户、拉取 Salesforce 数据、起草个性化邮件，早上留给你审批后发送。招聘筛选：自动浏览 LinkedIn 简历、匹配岗位需求、生成候选人短名单并安排面试。费用管理：从 Gmail 提取收据、分类归档、核对金额，定期生成报销报告。Bug 复现：自动登录测试环境、按步骤复现用户提交的 Bug、提交工单并移交给调试 Bot 跟进。内容运营：每小时扫描 Slack 频道查找会议信息，提醒相关人员并提供社交用语建议。# AI工具# AI项目和框架©版权声明本站文章版权归AI工具集所有，未经允许禁止任何形式的转载。上一篇ClawTeams - AI 电商运营协同平台，专家 Agent 并行执行下一篇dots.tts - 小红书联合上海交大开源的语音合成基座模型相关文章Agent Zero – 开源可扩展AI框架，通过用户指令和任务动态学习AI小集3Infinity – 字节跳动推出的高分辨率图像生成模型AI小集3ConsisID – 北大联合鹏城实验室等机构推出的文本到视频生成模型AI小集2SAM 2 – Meta推出的AI对象分割模型AI小集3Open-o3 Video – 北大联合字节开源的视频推理模型AI小集2Replymer – AI产品营销工具，用自然真实的方式推广产品AI小集3暂无评论再想想发表评论暂无评论…热门工具Loomy即梦AISekoAiPPT秘塔AI搜索妙呀堆友Agent美图设计室绘蛙AIupdream办公小浣熊秒哒最新收录OnSoloEpixa知漫剧QwenPawRenoiseLawbot脱敏猫最新文章秒悟Meoo – 1分钟生成前后端完整的网站，真0门槛开发！4小时前HiDream-O1-World – 智象未来推出的全模态交互式世界模型4小时前DeepSeek Harness Desktop – 开源的 DeepSeek Harness 桌面端4小时前dsh-TUI – 开源的 DeepSeek Harness 终端界面插件10小时前DSH Desktop – 开源的 DeepSeek Harness 本地桌面应用13小时前WorkSwarm – openJiuwen 推出的蜂群办公智能体1天前Happy Shrimp – 阿里 ATH 事业群推出的 AI 音乐平台1天前Copula Lab – AI 数据基础设施平台，提供专家级训练数据2天前Qwen3.8-27B – 阿里千问团队开源的新一代 AI 大模型2天前OnSolo – 腾讯推出的 AI 原生创作平台2天前MiniMax Music 3.0 – MiniMax 开源的音乐生成模型3天前dots3-note preview – 小红书开源的多模态 MoE 模型3天前GLM-5.3 – 智谱最新开源的基座大模型，编程能力提升50%3天前Gemini 3.7 Flash – Google DeepMind 推出最新的主力模型3天前wigolo – 开源的本地搜索工具，支持接入多种编程 Agent3天前

感谢IT之家网友shutiao、木木人工、Kazuto、Domado、浪花庆山粉、AlanMac、华南吴彦祖、LL J、乌蝇哥的左手、files、火瓶座的线索投递！IT之家8 月 14 日消息，当地时间 8 月 13 日，谷歌宣布推出 Gemini 3.7 Flash 模型，专为编程与 Agents 场景打造。IT之家注意到，本次发布距 Gemini 3.6 Flash 发布仅间隔约三周，谷歌称该版本的推出直接源于开发者的反馈以及一系列算法创新。Gemini 3.7 Flash 在软件工程、知识工作及网页开发工作流方面均实现了大幅提升。此外，在今年年底前（截至 2026 年 12 月 31 日），Gemini 3.7 Flash 将以首发优惠价提供（Gemini 3.6 Flash 也将有相同优惠），仅为原 3.6 Flash 每百万 Token 价格的一半：输入：每百万 Token 0.75 美元（现汇率约合 5.1 元人民币）输出：每百万 Token 3.75 美元（现汇率约合 25.3 元人民币）即日起，面向 160 多个国家和地区的 Google AI Pro 及 Ultra 订阅用户推出的 Gemini Spark 将正式接入 Gemini 3.7 Flash。本次模型升级显著提升了 Spark 处理知识工作的效率，优化了其对 Google Workspace 系列应用工具的使用能力，从而为复杂且跨技能的综合工作流提供更高的准确度与更高质量的输出。此外，Gemini 3.7 Flash 推出了全新的安全防御机制，在严防化学、生物、放射性及核（CBRN）领域和网络攻击滥用的同时，确保合法且有益的应用场景得以顺利实施。