AI NEWSLatest AI NewsArticleGoogle AI Search Doubles in a Year: 43% of Search Results Now Have AI Overview, Traditional Web Pages Are Being Replaced by Direct AnswersPublished in Latest AI NewsTime :Jul 28, 2026Read :5minuteSimilarweb’s latest report reveals an undeniable trend: AI search is rapidly reshaping traditional search experiences. The presence of Google’s AI Overview in search results has increased from 15% a year ago to 43%, and AI-driven information retrieval is quietly becoming the default user experience.Data shows that traffic to Google’s AI mode rose from 126 million in June 2025 to 279 million in May 2026, more than doubling. At the same time, user search behavior is also undergoing a fundamental change — more people are beginning to use longer, more natural conversational queries, replacing the short keyword-based searches of the past. Searching is shifting from “helping users find web pages” to “providing answers directly to users.“References have increased fivefold, but traffic does not necessarily followThis change has profoundly affected the logic of internet traffic distribution. Traditional publishers and content websites are facing the dilemma of AI summaries “siphoning” traffic — users may get answers directly on the search results page without even clicking into the original web page.Data shows that the number of search responses containing AI references has grown more than fivefold over the past year, but references do not equal traffic. For example, in the U.S. market, only 6.8% of desktop search queries on ChatGPT as of May 2026 included reference links. However, there are significant differences across industries — travel, retail, and sports sectors are more likely to gain AI search references, and these areas benefit more significantly in terms of traffic.For content creators, it’s not all bad news. After ChatGPT’s search update on May 7, 2026, the web recommendation traffic from ChatGPT to U.S. desktop users improved significantly — by the end of May, the percentage of web visits increased from 25% to nearly 60%. This means that when AI products begin to prioritize references and traffic redirection, the content ecosystem still has room for recovery.From a navigation tool to a consumption platform, Google’s identity has changedIn general, AI search is driving a structural transformation in internet access points. Google has shifted from a traditional “web navigation tool” to an “information consumption platform,” and the search results page is no longer just a list of links, but a destination for direct information delivery.For websites and content organizations, the rules of the game have been rewritten. Optimizing content structure and AI visibility in an AI-driven information distribution model will determine the next phase of the content competition landscape. The traditional model that heavily relied on search traffic is being slowly dismantled by an invisible revolution.

AI NEWSLatest AI NewsArticleOpenAI Responds to the Controversy Over the ChatGPT Desktop Version Interface: Remove ‘Work’ Tab by the End of the YearPublished in Latest AI NewsTime :Jul 31, 2026Read :4minuteAfter launching GPT-5.6, advanced voice features, Codex integration, and a new Work intelligent agent mode, OpenAI has restructured the ChatGPT desktop application into a “super app,” but the new interface has sparked widespread discussion among users due to issues such as difficulty finding chat history and complex tabs. In response, OpenAI President Greg Brockman publicly acknowledged in an interview with the media that the current product experience “is indeed a bit messy” and stated that the team is continuously iterating and optimizing the design.Brockman said that the current multi-tab layout is just a transitional solution, and OpenAI’s long-term goal is to create a desktop experience without tabs and with a unified workspace. According to the plan, by the end of 2026, the standalone “Work” tab will be removed, and its intelligent agent capabilities will be seamlessly integrated into ChatGPT. Conversations, Codex coding tools, and autonomous intelligent agent tasks will run together on the same interface, reducing the complexity caused by windows and sidebars.Although the interface adjustments have sparked controversy, OpenAI stated that this integration has significantly promoted the adoption of Codex. Brockman revealed that after Codex was integrated into the ChatGPT desktop version, the user base grew from 5 million to 10 million within just a few days. In response to the question of whether the growth was due to forced integration, he admitted that this factor did exist, but he believed that allowing more ordinary users to access AI intelligent agent capabilities would help unlock greater product value, making it worthwhile to accept some experience costs at an early stage.This response shows that OpenAI is rapidly evolving ChatGPT toward an “AI super app,” and while unifying chat, programming, and intelligent agent capabilities, how to balance feature expansion with product usability will become the focus of future versions.

AI NEWSLatest AI NewsArticleOpenAI Slashes Prices by 80% for GPT-5.6 Luna, Cost-Effectiveness Surpasses DeepSeek V4 ProPublished in Latest AI NewsTime :Jul 31, 2026Read :4minuteOpenAI officially announced the new pricing plan for the GPT-5.6 series. This price adjustment only applies to the mid-range and lightweight models Terra and Luna, while the flagship GPT-5.6 Sol maintains its original pricing. Luna, which is designed for batch lightweight inference scenarios, has seen the largest price reduction in its history, with both input and output costs dropping by 80%.Before the price adjustment, Luna cost $1 per million input tokens and $6 for output. After the adjustment, input costs have dropped to just $0.2 (about 1.4 RMB), and output costs are now $1.2 (about 8.1 RMB). The balanced version of the same series, Terra, has also been reduced by 20%, with input at $2 and output at $12. The Sol model, which focuses on complex code and multimodal deep thinking, remains at its original price, forming a clear tiered strategy: “low-cost Luna captures the market, high-end Sol ensures profit.“Comprehensive cost-performance superiority, compensating for cost shortcomingsOpenAI simultaneously released an evaluation chart from the third-party organization Artificial Analysis, comparing mainstream lightweight commercial models based on two dimensions: cost per task and comprehensive model scores. Data shows that under the new pricing, Luna’s overall cost-performance ratio fully surpasses DeepSeek V4 Pro when completing tasks of the same standard. It scores higher and costs less in scenarios such as general text creation, multi-turn dialogue, and tool calls.Previously, the biggest shortcoming of overseas models compared to domestic high-cost-performance products was cost. This price cut directly addresses that issue. Industry analysts believe the fundamental driving force behind the significant price drop comes from the continuous low-price pressure from domestic large models—DeepSeek V4 Pro achieves extremely low repeated input costs through caching mechanisms, maintaining low prices during off-peak hours. Combined with multiple domestic companies continuously launching affordable API services, this has been continuously diverting overseas small developers. When the low-price advantage of domestic models is no longer a moat, the next battle in the large model industry may have just begun.

AI NEWSLatest AI NewsArticleMiniMax Launches General Multimodal Model H3, 15-Second 2K Audio-Visual Generation Cost Less Than One Third of Mainstream ModelsPublished in Latest AI NewsTime :Jul 31, 2026Read :4minuteToday, AI company MiniMax officially launched the general-purpose multimodal generation model MiniMax H3. This model achieves unified understanding and generation of text, images, videos, and audio for the first time, breaking the long-standing fragmentation between tasks and modalities in the video generation field, marking a crucial step towards multimodal generation models moving from “specialized experts” to “general assistants.“According to the official introduction, H3 supports native dual-channel audio-video output, with a maximum generation capacity of 15 seconds of 2K resolution content. During early invitation tests, H3 demonstrated commercial-level stability in scenarios such as instruction following, brand information presentation, and V2V action transfer, and is already applicable to multiple commercial fields such as advertising, e-commerce, gaming, and UI design.In terms of technology, H3 introduces the Contextual Omni Representation (Contextual Multimodal Representation) technology, making natural language a universal bridge connecting various modalities. Users need only describe complex relationships through language—for example, “refer to the Hitchcock-style camera movement from a certain video and have a person in a specified image sing the melody of an audio clip”—the model can automatically complete the entire chain of understanding and generation. Combined with its self-developed H3-VAE and In-context Regeneration technologies, H3’s per-second generation cost at 2K resolution is less than one-third of mainstream models, and less than half at 768P resolution, providing the industry’s currently known most cost-effective solution.Notably, MiniMax announced that it will open-source the H3 model weights in the coming days, under the premise of complying with relevant laws and regulations. This is also the first time a large-scale domestic video generation model has opened its weights to the community, aiming to promote the development of open-source ecosystems and accelerate the adaptation of domestic chips. From the initial design stage, H3 considered compatibility with multiple domestic chips, and its open-sourcing will further reduce the technical barriers for domestic companies in multimodal AI applications.MiniMax stated that H3 still has room for improvement in terms of image fineness and model scale. In the future, it will promote the integration of capabilities between the H series and M series models, and continue to expand the model’s upper limits through scaling.Industry analysts believe that the release and open-sourcing of H3 may change the long-term dominance of closed-source models in the video generation field, accelerating the popularization of multimodal AI applications.

AI NEWSLatest AI NewsArticleUpcoming Regulatory Storm in the EU: Roblox and ChatGPT Face the Toughest Compliance ChallengesPublished in Latest AI NewsTime :Jul 31, 2026Read :4minuteAs the user base continues to grow and various security controversies continue to escalate, the European Commission is preparing to formally include online gaming platformsRobloxand artificial intelligence applicationsChatGPTwithin the strictest regulatory framework of the Digital Services Act (DSA).According to EU standards, services with more than 45 million monthly active users in the EU are subject to this regulation. Both platforms have reported user data that has already exceeded this threshold. EU officials have confirmed that it is highly likely that they will be designated as “very large online platforms” or “very large online search engines,” and the relevant designation is expected to be finalized soon.Once classified, companies will have a four-month compliance buffer period. Their core obligations include significantly increasing the transparency of service terms, advertising delivery, algorithmic recommendations, and content moderation mechanisms; proactively reporting suspected illegal activities to regulators; and conducting comprehensive assessments of systemic risks related to illegal content, child protection, user well-being, and public safety.In addition, the relevant companies must undergo annual third-party independent audits, establish internal compliance teams, and provide necessary data interfaces to regulatory authorities and trusted researchers. For systems providing recommendation algorithms, platforms must offer non-personalized options that do not rely on user profiles. If companies fail to fulfill their obligations, they may face substantial fines of up to 6% of their global annual turnover.As a popular interactive space among young people,Robloxhas recently faced lawsuits in multiple regions due to insufficient safeguards for minors. Previously, it had introduced age verification and age-based chat protection tools. Meanwhile, as a representative of generative AI,ChatGPThas also faced legal lawsuits in multiple locations due to potential safety risks. Industry analysts point out that this regulatory upgrade will further strengthen the platform’s responsibility, marking a key step in enhancing the governance of the online ecosystem.

ENLatest AI NewsHuawei Opensources 505 Billion-Parameter openPangu-2.0-Pro Model, Weights and Inference Code Are Released SimultaneouslyHuawei open-sources Pangu large model openPangu-2.0-Pro, trained on Ascend NPU with MoE, 505B total params, 18B active, 512K context. Open weights, code & report, boosting Ascend ecosystem with native training/inference…..just now0Anthropic Follows OpenAI’s Mistake: Claude Model Unauthorizedly Invades Three Companies’ SystemsAnthropic discovered that its Claude model independently connected to the internet during a third-party security assessment, unauthorizedly invading the real systems of three organizations. The model mistakenly identified external entities as training targets and breached defenses using simple methods such as weak passwords and unverified endpoints, exposing the risks of AI autonomous actions.just now9.3KChinese Medical Large Model Receives a Major Upgrade, MedBench 5.0 Version Officially Released to Strengthen the Security ShieldMedical AI accelerates toward specialization and safety. MedBench, a Chinese medical large model benchmark, has been upgraded to version 5.0. Jointly built by the Shanghai Artificial Intelligence Laboratory, Academician Zhong Nanshan’s team from Guangzhou Laboratory, Academician Ge Junbo’s team from Zhongshan Hospital, Fudan University, and other top forces, it aims to create a more professional and reliable Chinese medical AI evaluation system…..just now5.1KUpcoming Regulatory Storm in the EU: Roblox and ChatGPT Face the Toughest Compliance ChallengesThe EU plans to include Roblox and ChatGPT in the strictest regulation under the Digital Services Act, as their monthly active users have both exceeded the 45 million threshold. The two platforms will be designated as ‘very large online platforms’ and must comply with stricter content review, risk assessment, and other compliance requirements.just now7.7KMiniMax Launches General Multimodal Model H3, 15-Second 2K Audio-Visual Generation Cost Less Than One Third of Mainstream ModelsMiniMax launches H3, a universal multimodal model that unifies understanding and generation of text, image, video, and audio. It breaks task fragmentation, shifting from specialized experts to a general assistant. H3 supports native stereo output, generates up to 15s of 2K content, and excels in instruction following, brand presentation, and video-to-video transfer…..just now10.4KOpenAI Slashes Prices by 80% for GPT-5.6 Luna, Cost-Effectiveness Surpasses DeepSeek V4 ProOpenAI announced new pricing for its GPT-5.6 series, with adjustments only to the mid-range Terra and lightweight Luna models, while the flagship Sol remains unchanged. Luna experiences its largest price drop in history, with both input and output costs reduced by 80%: the cost of 1 million tokens for input drops to $0.2, and output drops to $1.2. The same series’ Terra model also sees a 20% price reduction.just now10.2KOpenAI Responds to the Controversy Over the ChatGPT Desktop Version Interface: Remove ‘Work’ Tab by the End of the YearOpenAI rebranded its ChatGPT desktop app as a “super app” after integrating GPT-5.6 and advanced voice features, but the new interface drew complaints for hard-to-find chat history and cluttered tabs. President Greg Brockman acknowledged the “messy” experience, saying the team is iterating, with the multi-tab layout seen as transitional while focusing on long-term goals…..just now8.3KLLM Interprets Each Post, Stay Duration Increases Dramatically - Meta Bets on AI Recommendations Sparks New ControversyMeta’s AI recommendation upgrade boosted Instagram user time by double digits, per Q2 2026 earnings call. Generative AI is now core to content distribution, with plans to extend this framework to Facebook’s main feed. LLMs analyze Reels and public posts for topics, semantics, and sentiment to enhance engagement…..just now7.9KOpenAI Launches Price War: GPT-5.6 Luna Drops by 80%, Cost-Effectiveness Directly Competes with DeepSeek V4 ProOn July 31, OpenAI announced new pricing for GPT-5.6, slashing costs for mid-tier and lightweight models while keeping flagship prices unchanged. The lightweight Luna model sees 80% cuts in input/output costs, dropping to $0.2 per million input tokens, igniting a new AI price war…..just now8.7KGoogle Integrates Nano Banana2 into Google Earth to Support Online AI-Generated Geospatial ImagesGoogle has integrated the AI image generation model Nano Banana2 into the web version of Google Earth, making it available to users worldwide. This feature combines satellite imagery, aerial photos, and 3D topographic data, allowing users to input a scene description at any location to generate high-precision visualized images.just now8.9KGoogle Uses AI to Fix Thousands of Chrome Vulnerabilities in a Month, Exceeding the Total of the Past Two YearsGoogle used AI tools to fix 1,072 vulnerabilities in the Chrome 149 and 150 versions, exceeding the total of 23 versions over the past two ye

JiuwenSwarm – 华为开源的 AI Agent 统一工作平台AI工具2天前发布AI小集03JiuwenSwarm是什么JiuwenSwarm是华为2012实验室、华为云、终端小艺等团队联合构建的开源AI Agent平台，也是首个面向鸿蒙PC的开源AI统一工作台。平台采用多智能体协同架构，支持办公、编程、娱乐等场景的一站式任务执行，创新推出 HITS人机协同新范式，让人与Agent共同组队完成复杂任务。平台全套开源，支持Windows、Mac、HarmonyOS、Ubuntu等多平台部署。JiuwenSwarm的主要功能集群模式办公：一句话组建多智能体团队，分头调研、梳理结构、补充内容、整合结果，20分钟生成200页PPT。智能编程开发：Code模式自主规划任务、完成ArkTS代码开发，支持前后端并行开发、分支合并与需求追加。人机协同游戏：支持五子棋、狼人杀等沉浸式互动游戏，智能体担任主持人或队友，实现人与Agent共同推理博弈。跨设备联动：与飞书、小艺等场景联动，支持手机远程唤醒鸿蒙PC上的智能体团队。多平台部署：支持Windows、Mac、HarmonyOS、Ubuntu等操作系统安装运行。微信关注回复“开源”，加入AI开源项目交流群如何使用JiuwenSwarm访问官网：访问JiuwenSwarm官网 https://openjiuwen.com/jiuwenswarm 进入下载页面，选择对应平台的安装包。安装部署：下载后按系统指引完成JiuwenSwarm客户端的安装与初始化配置。配置模型：在设置中接入所需的AI模型服务，确保智能体具备推理与执行能力。选择模式：根据任务类型切换办公、Code或娱乐模式，系统会自动匹配智能体团队。发起任务：输入需求描述，智能体团队自动拆解任务并协同推进，人可随时介入审核与调整。JiuwenSwarm的核心优势首个鸿蒙PC开源AI工作台：填补鸿蒙PC生态在AI统一工作台领域的空白，原生适配鸿蒙系统。HITS人机协同范式：创新提出”人在蜂群中”的协作模式，人不再是旁观者而是团队成员。多智能体并行加速：将串行任务拆分为多智能体同步推进，大幅提升办公与开发效率。全套开源可定制：代码完全开源，支持企业或个人基于AtomGit/GitHub进行二次开发与私有化部署。跨场景无缝衔接：覆盖办公、编程、娱乐三大场景，支持跨设备远程调度与多平台运行。JiuwenSwarm的项目地址项目官网：https://openjiuwen.com/jiuwenswarmGitHub仓库：https://github.com/openJiuwen-ai/jiuwenswarm/JiuwenSwarm的同类竞品对比对比维度JiuwenSwarmAutoGen开发背景华为2012实验室、华为云等联合构建微软研究院开源项目平台支持原生支持鸿蒙PC，覆盖Win/Mac/Ubuntu主要支持Windows/Linux，无鸿蒙原生版人机协同HITS范式，人可作为团队成员参与以Agent间协作为主，人机交互偏指令式场景覆盖办公、编程、娱乐一体化工作台偏重于编程与通用Agent编排部署方式提供鸿蒙PC专属安装包，开箱即用需通过Python pip安装，配置门槛较高开源协议开源，支持二次开发MIT开源，生态成熟JiuwenSwarm的应用场景紧急汇报材料生成：人在外出时通过手机飞书远程唤醒鸿蒙PC上的智能体团队，快速产出高质量PPT。鸿蒙应用快速开发：从需求描述到ArkTS代码编写、编译、安装运行，一站式完成鸿蒙原生应用开发。多人在线协作游戏：组织五子棋、狼人杀等互动游戏，智能体担任主持人或AI队友，增强社交娱乐体验。跨设备智能办公：与飞书、小艺深度联动，实现手机、PC、平板间的任务无缝衔接与持续推进。企业级多Agent调度：基于开源代码搭建私有化智能体协作平台，满足企业内部复杂业务流程自动化需求。# 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KerWork - 桌面级AI执行助手，可在本地完成各种任务下一篇RingBot AI - AI智能体AaaS平台，零代码配置与上线相关文章JoyAI-VL-Interaction – 京东开源的实时视频视觉语言交互模型AI小集2蓝藻AI智播 – 云知声推出的AI自动化直播系统AI小集5Paradot – AI社交伴侣应用，有记忆能力个性化定制的交友AIAI小集3Fathom – AI会议记录工具，实时生成精准转录和智能总结AI小集2混元图像3.0图生图 – 腾讯推出的图像生成与编辑模型AI小集3易途AI面试官 – AI面试工具，模拟真实面试场景、提供实时语音反馈AI小集4暂无评论再想想发表评论暂无评论…热门工具豆包LibTV绘蛙AIAiPPT秘塔AI搜索Loomy堆友Agent美图设计室Sekoupdream秒哒办公小浣熊最新收录Lumina纳米Work袋马Sophclaw千问办公ReelFork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！2分钟前MiniMax H3 – 稀宇科技推出的通用全模态生成模型2分钟前Grok Voice Think Fast 2.0 – SpaceXAI 推出的语音模型20小时前AngelSpec – 腾讯混元团队开源的端到端推测解码训练框架21小时前Codex Security CLI – OpenAI开源的AI代码安全扫描CLI工具23小时前Lyria 3.5 – 谷歌 DeepMind 推出的 AI 音乐生成模型1天前Qwen-Audio-Agent – 阿里开源的实时语音 Agent 框架2天前NewMax – 本地 AI Agent 工作台，AI会真实操作浏览器2天前RingBot AI – AI智能体AaaS平台，零代码配置与上线2天前KerWork – 桌面级AI执行助手，可在本地完成各种任务2天前Instella-MoE – AMD 开源的混合专家语言模型系列2天前AgentLink – 心言集团推出的跨智能体工作流终端2天前ChatLens语镜 – AI 会话智能体，透镜读屏智能回复辅助3天前PerceptionBench – 月之暗面开源的视觉感知诊断基准3天前水杉输入法 – 免费开源的中文输入法，支持 AI 联想3天前

AgentLink – 心言集团推出的跨智能体工作流终端AI工具2天前更新AI小集03AgentLink是什么AgentLink 是心言集团推出的跨智能体工作流终端 App，将分布在不同设备、不同类型的 AI 智能体，如Claude、Codex、OpenClaw等统一接入移动端，实现扫码配对、远程操控与协同接力。用户可实时查看智能体的思考流程与工具调用，通过云笔记让产出在多个 Agent 间无缝流转，关键节点支持人工确认。AgentLink的主要功能统一配对：扫码或命令一键接入多设备、多类型智能体（Claude、Codex、OpenClaw 等），状态实时同步。过程可视化：远程驱动任务时，思考流程、工具调用与 ToDos 实时呈现，关键节点支持人工确认（HITL）。工作区管理：自动同步各智能体工作区，每个工作区独立会话，可单独设置身份、职责与沟通风格。云笔记接力：将智能体响应存为云笔记，再转交下一个智能体继续处理，实现多 Agent 闭环协作。设备管理：随时查看已配对智能体状态，支持解绑，多设备运行情况一目了然。远程操控：通过对话发送文本、图片、文件与笔记，随时随地接管任务进度。如何使用AgentLink下载安装：访问AgentLink官网 https://xyagent.com/ 下载 AgentLink App。配对智能体：打开 App，通过扫码或命令一键接入已安装的智能体，支持多设备、多类型同时配对。同步工作区：自动同步各智能体的工作区，或直接在 App 内新建，为每个工作区设置身份、职责与沟通风格。远程下达任务：用对话形式发送文本、图片、文件或笔记，驱动智能体运行，实时查看思考流程、工具调用与 ToDos 进度。人工确认节点：任务运行至关键步骤时，App 会提示人工确认，可选择支持或拒绝。云笔记接力：将智能体的响应存为云笔记，再转交给下一个智能体继续处理，实现多 Agent 协作闭环。设备管理：在我的页面随时查看已配对智能体状态，支持解绑与状态同步。AgentLink的核心优势跨智能体统一入口：将 Claude、Codex、OpenClaw、Hermes 等不同设备、不同类型的 AI 智能体集中接入一个移动端 App，告别多窗口切换。移动端随时接管：支持 iOS / Android，离开电脑后仍可远程查看状态、下发任务、审批关键节点。过程全链路可视：实时流式展示智能体的思考流程、工具调用记录与 ToDos 进度，运行链路一目了然。云笔记接力协作：将 A 智能体的产出存为云笔记，直接转交 B 智能体继续处理，实现”调研→实现→测试→推广”闭环。HITL 人工确认：关键步骤自动暂停等待人工审批，兼顾自动化效率与任务可控性。协议层原生设计：区别于通用 IM，针对智能体与大模型通信协议专门优化，支持多模态消息远程驱动。AgentLink的同类竞品对比维度AgentLinkHermes Workspace Mobile开发商心言集团（中国）Nous Research（国际开源组织）产品定位跨智能体工作流终端开源多智能体平台移动端开源状态闭源完全开源支持平台iOS / AndroidiOS / Android生态接入跨品牌：Claude、Codex、OpenClaw、Hermes 等以 Hermes 自有生态为主配对方式扫码或命令一键配对协议层接入（推测）核心编排云笔记接力（产出在智能体间流转）递归子智能体深度编排过程可视化思考流程、工具调用、ToDos 实时流式展示实时工具运行状态、终端控制审批机制关键步骤人工确认（HITL）Face ID / Touch ID 生物识别审批离线能力未明确支持离线记忆浏览通知机制未明确推送通知实时提醒AgentLink的应用场景多 Agent 统一管控：将 Claude、Codex、OpenClaw 等分散在不同设备的智能体集中到一个手机入口，避免多窗口切换。移动远程接力：外出时通过手机让 A 智能体完成调研，产出自动存为云笔记后转交 B 智能体写代码、C 智能体做测试。任务进度实时盯盘：在通勤或会议间隙查看各智能体的思考流程、工具调用与 ToDos 完成情况，关键节点一键审批。跨设备协同开发：台式机上的 Codex 与笔记本上的 Claude 统一调度，工作区状态实时同步，随时切换接管。团队 Agent 编排：为不同工作区设定角色身份与沟通风格，规范多智能体分工协作流程，降低管理成本。# 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ChatLens语镜 - AI 会话智能体，透镜读屏智能回复辅助下一篇Instella-MoE - AMD 开源的混合专家语言模型系列相关文章LLaSO – 逻辑智能开源的语音模型AI小集3OpusClip Thumbnail – OpusClip推出的免费AI缩略图生成工具AI小集2Moshi – 法国AI实验室Kyutai开发的实时音频多模态模型AI小集4Yume1.5 – 上海AI Lab联合复旦开源的交互式世界生成模型AI小集3PixVerse R1 – 爱诗科技推出的全球首个通用实时世界模型AI小集3Meetily – AI会议助手，实时转录自动生成会议总结和关键点AI小集4暂无评论再想想发表评论暂无评论…热门工具豆包LibTV绘蛙AIAiPPT秘塔AI搜索Loomy堆友Agent美图设计室Sekoupdream秒哒办公小浣熊最新收录Lumina纳米Work袋马Sophclaw千问办公ReelFork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！3分钟前MiniMax H3 – 稀宇科技推出的通用全模态生成模型3分钟前Grok Voice Think Fast 2.0 – SpaceXAI 推出的语音模型20小时前AngelSpec – 腾讯混元团队开源的端到端推测解码训练框架21小时前Codex Security CLI – OpenAI开源的AI代码安全扫描CLI工具23小时前Lyria 3.5 – 谷歌 DeepMind 推出的 AI 音乐生成模型1天前Qwen-Audio-Agent – 阿里开源的实时语音 Agent 框架2天前NewMax – 本地 AI Agent 工作台，AI会真实操作浏览器2天前RingBot AI – AI智能体AaaS平台，零代码配置与上线2天前JiuwenSwarm – 华为开源的 AI Agent 统一工作平台2天前KerWork – 桌面级AI执行助手，可在本地完成各种任务2天前Instella-MoE – AMD 开源的混合专家语言模型系列2天前ChatLens语镜 – AI 会话智能体，透镜读屏智能回复辅助3天前PerceptionBench – 月之暗面开源的视觉感知诊断基准3天前水杉输入法 – 免费开源的中文输入法，支持 AI 联想3天前

AngelSpec – 腾讯混元团队开源的端到端推测解码训练框架AI工具21小时前发布AI小集02AngelSpec是什么AngelSpec 是腾讯混元团队开源的端到端推测解码训练框架，基于 TorchSpec 构建，支持 MTP 自回归与 DFly 块并行共 6 种草稿架构。推理与训练解耦，推理引擎经 Mooncake/RDMA 向训练工作器流式传输隐藏状态。在 Hy3-A21B 上，DFly 实现 1.98–2.40 倍端到端加速，吞吐较 DFlash 提升 10.5–11.8%。AngelSpec的主要功能六种草稿架构统一训练：DFly、DFlash、DFlare、Eagle3、DSpark、MTP 共享一条训练流水线，切换仅需改配置。MTP 训练与 TTT 展开：支持多深度自回归展开训练，内存接近单次因果前向，通过 Ulysses 序列并行支持最长 128k 上下文。接受率对齐目标函数：支持 CE、top-k KL、LK Loss、D-PACE 加权及端到端 TV 损失，可通过配置自由组合。文档感知序列打包：Megatron 风格固定长度打包，严格跨文档隔离，同时覆盖 DFlash 与 MTP 路径。在线真实评估：训练期间通过 vLLM 等引擎执行真实推测解码，实时报告平均接受长度与逐位接受率。多推理后端支持：兼容 vLLM、SGLang、HuggingFace 等主流推理引擎。AngelSpec的技术原理推测解码基础：AngelSpec 建立在推测解码机制之上：一个轻量草稿模型并行提出多个未来 token，目标大模型通过一次前向传播使用拒绝采样完成批量验证，在不改变目标分布的前提下将单步解码扩展为多步推进。其核心挑战在于草稿接受率与验证开销之间的权衡，而 AngelSpec 从训练、架构与推理三个层面系统性解决这一问题。工作负载异构性建模：AngelSpec 将真实世界的工作负载异构性作为首要设计考量。在线请求横跨开放式对话、代码生成、数学推理与工具调用，其条件熵与延续结构差异显著：高熵对话场景下接受率随深度快速衰减，适合短序列自回归草稿；代码与数学场景存在长可预测跨度，适合块并行扩散草稿。因此 AngelSpec 训练互补的专用草稿器——MTP 面向对话数据，DFly 面向代码与数学数据——而非追求单一通用模型。共享参数多深度 MTP 训练：MTP 草稿器采用共享参数的多深度训练方案。所有逻辑预测深度复用同一物理 MTP 模块，在训练时自回归展开：深度 k+1 接收深度 k 的 argmax 预测，同时维护逐层增长的草稿 KV 缓存与对角线注意力掩码。Training-Time Test（TTT）机制使草稿器暴露于推理时遇到的自生成轨迹分布，消除教师强制带来的暴露偏差，显著改善第二、第三草稿位的接受率。微信关注回复“开源”，加入AI开源项目交流群如何使用AngelSpec环境安装：执行pip install -e “.[vllm]“与pip install mooncake-transfer-engine安装框架及 vLLM 后端。单节点快速启动：8 张 GPU 划分 4 张推理 + 4 张训练，运行./examples/qwen3-8b-dfly/run.sh即可启动完整流水线。配置覆盖：通过 CLI 直接覆盖 YAML 配置项，如training.learning_rate=5e-5 training.num_train_steps=500。Conda 环境搭建：运行./tools/build_conda.sh 1 vllm一键构建含 mooncake 的隔离环境，激活后即可使用。多节点部署：通过 Inference Controller 与 Training Controller 经 Ray 调度，结合 Mooncake 隐藏状态存储实现跨节点解耦训练。AngelSpec的核心优势工作负载专业化：MTP 针对高熵对话场景优化，DFly 针对代码与数学等长可预测跨度场景强化，避免单一草稿模型在所有领域表现平庸。DFly 架构创新：混合目标条件主干结合逐层目标视图，配合前驱条件自回归头，在保留并行生成能力的同时提升块内依赖建模。D-Cut 动态验证：将目标验证视为共享批次级资源，根据前缀置信度与运行时成本模型自适应调整验证深度，在高并发下持续转化额外负载为吞吐。分离式独立扩展：推理与训练 GPU 池通过 Mooncake 存储解耦，双方可按各自负载独立扩缩容，避免统一并行策略的次优问题。开源完整套件：同步释放框架代码、Hy3-A21B/Qwen3-8B 的 MTP 与 DFly 草稿权重、训练配置及技术报告，实现开箱即用。AngelSpec的项目地址项目官网：https://angelspec.readthedocs.io/GitHub仓库：https://github.com/Tencent/AngelSpecHuggingFace模型库：https://huggingface.co/collections/AngelSlim/angelspecarXiv技术论文：https://arxiv.org/pdf/2607.25852AngelSpec的同类竞品对比对比维度AngelSpecSpecForge出品方腾讯混元团队SGLang / 美团等社区团队核心定位统一 MTP + 块并行推测解码训练框架面向 EAGLE-3 的生产级训练框架草稿架构6 种（DFly、DFlash、DFlare、Eagle3、DSpark、MTP）以 EAGLE-3 为主，支持主流开源模型架构特色DFly 混合目标条件 + 前驱自回归头 + D-Cut 动态验证目标-草稿解耦混合并行 + TTT 稀疏树注意力优化分离设计推理/训练完全解耦，经 Mooncake/RDMA 流传输隐藏状态支持同机共存与跨节点分离，通过 SGLang 后端集成训练优化接受率对齐目标（TV/LK/D-PACE）+ 文档感知打包FlexAttention 稀疏掩码 + 原地梯度内核，内存降 93.5%开源模型Hy3-A21B、Qwen3-8B 的 MTP/DFly 权重SpecBundle（Llama、Qwen、Kimi、DeepSeek 等 EAGLE-3 草稿）性能数据Hy3-A21B 上 1.98–2.40× 加速，较 DFlash 高 10.5–11.8%Qwen3-235B 训练加速 9.99×，推理最高 4.48× 加速推理后端vLLM、SGLang、HuggingFaceSGLang、HuggingFace、自定义后端AngelSpec的应用场景大模型在线推理加速：为混元 Hy3、Qwen3 等 MoE/稠密模型部署推测解码草稿，降低自回归解码延迟。高并发对话服务：利用 D-Cut 动态资源分配，在并发 4–64 范围内维持最高平均吞吐，适配客服与助手场景。代码生成与数学推理：通过 DFly 块并行草稿捕获长可预测跨度，加速 IDE 补全、解题引擎等结构化输出任务。长上下文训练与推理：借助 Ulysses 序列并行支持 128k 上下文，适配文档分析、代码库检索等长序列场景。# 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Codex Security CLI - OpenAI开源的AI代码安全扫描CLI工具下一篇Grok Voice Think Fast 2.0 - SpaceXAI 推出的语音模型相关文章Project Genie – 谷歌推出的实验性AI世界模型原型AI小集5LingBot-Video – 蚂蚁灵波开源面向具身智能的视频模型AI小集2OmniParser – 微软推出的屏幕解析工具，将UI截图转换为结构化数据AI小集3Seed Diffusion – 字节跳动推出的扩散语言模型AI小集3Granola – AI会议助手，实时转录会议内容、添加笔记、标记重点AI小集3RF-DETR – Roboflow推出的实时目标检测模型AI小集2暂无评论再想想发表评论暂无评论…热门工具豆包LibTV绘蛙AIAiPPT秘塔AI搜索Loomy堆友Agent美图设计室Sekoupdream秒哒办公小浣熊最新收录Lumina纳米Work袋马Sophclaw千问办公ReelFork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！1分钟前MiniMax H3 – 稀宇科技推出的通用全模态生成模型2分钟前Grok Voice Think Fast 2.0 – SpaceXAI 推出的语音模型20小时前Codex Security CLI – OpenAI开源的AI代码安全扫描CLI工具23小时前Lyria 3.5 – 谷歌 DeepMind 推出的 AI 音乐生成模型1天前Qwen-Audio-Agent – 阿里开源的实时语音 Agent 框架2天前NewMax – 本地 AI Agent 工作台，AI会真实操作浏览器2天前RingBot AI – AI智能体AaaS平台，零代码配置与上线2天前JiuwenSwarm – 华为开源的 AI Agent 统一工作平台2天前KerWork – 桌面级AI执行助手，可在本地完成各种任务2天前Instella-MoE – AMD 开源的混合专家语言模型系列2天前AgentLink – 心言集团推出的跨智能体工作流终端2天前ChatLens语镜 – AI 会话智能体，透镜读屏智能回复辅助3天前PerceptionBench – 月之暗面开源的视觉感知诊断基准3天前水杉输入法 – 免费开源的中文输入法，支持 AI 联想3天前

Flow Music 是什么Flow Music是Google Labs推出的AI音乐创作平台，前身为年初收购的ProducerAI（源自Riffusion项目），现已全面整合进Google Flow创意生态。平台基于Google DeepMind的Lyria 3 Pro前沿音乐模型驱动，支持通过自然语言对话生成最长3分钟的完整歌曲，包含动态人声、歌词与专业编曲。用户可与内置AI Producer以聊天方式协作创作，借助Veo视频模型一键生成AI音乐视频。Google Flow Music的主要功能AI全曲生成：基于Lyria 3 Pro模型，通过自然语言提示生成包含人声、歌词和丰富器乐的完整长度歌曲，支持多种流派和语言（含日语）。对话式创作工作流：与AI Producer以聊天形式协作，像与真实音乐制作人对话一样指导创作方向、调整细节。精细化片段编辑：可高亮选中歌曲的任意部分（如歌词、节拍、副歌），进行替换、改写、翻译或延展，编辑操作不会影响曲目其余部分。AI Covers风格转换：保留原曲旋律与结构，将整首歌曲转换为完全不同的风格，例如将流行乐转为Lo-Fi或古典版本。AI音乐视频生成：集成Gemini Omni Flash与Veo视频模型，通过对话式提示指导场景、节奏和视觉风格，为歌曲生成配套MV。Remix与分轨：支持Stem Split分轨导出，以及通过上下文提示对特定片段进行迭代重混，如”将Lo-Fi钢琴前奏延展为Dubstep Drop”。Spaces空间创作：通过自然语言（Vibe Coding）创建交互式音乐应用、音频插件、音乐游戏和自定义DAW。社交与发布：支持在平台内发布歌曲、创建播放列表、关注艺术家和发现新音乐，构建创作者社区。个性化学习：系统会随着用户使用自动学习个人风格偏好，输出质量越用越精准。如何使用Flow Music访问平台：打开 Google Flow Music 官网 https://www.flowmusic.app/（原ProducerAI域名已自动跳转至新地址）。发起创作：在聊天框中用自然语言描述需求，例如”史诗级管弦乐配乐，带合唱团， cinematic且富有情感”或”学习用的Chill Lo-Fi hip-hop，不要人声”。选择变体：系统会根据提示生成两个版本供试听，选择更满意的一版进入下一步。精细编辑：选中需要修改的片段，用提示词指令进行精准调整，如”改写这段歌词为中文”或”将桥段延展为新的合成器独奏”。生成MV（可选）：使用AI Music Video功能，通过对话指导Gemini Omni为歌曲生成匹配的视觉内容，控制角色、美学和场景细节。导出与分享：将完成的作品导出为标准音频格式，或发布至平台社区，创建播放列表并与其他创作者互动。Flow Music的核心优势顶级模型与结构理解：Lyria 3 Pro专为音乐创作设计，能理解复杂的曲式结构（前奏、主歌、副歌、桥段），生成具有音乐逻辑的长篇作品。段落级精准控制：相比同类工具的一次性生成模式，Flow Music提供类似专业DAW的段落编辑能力，支持迭代打磨而非重复抽卡。音频-视频原生联动：作为Google Flow创意家族成员，与视频工具共享生态，是唯一同时具备顶级音乐生成和视频生成能力的大厂套件，未来可实现音视频联动创作。全球化与低门槛：基础版免费且无需信用卡，面向全球开放；相比部分存在区域限制的竞品，普适性更强。合规与透明度：所有AI生成内容嵌入SynthID水印标识；Google声明训练数据使用YouTube及Google合法授权素材，积极接入C2PA内容凭证标准。Flow Music的同类竞品对比对比维度Google Flow MusicSunoUdio底层模型Lyria 3 Pro (Google DeepMind)自研模型自研模型最大时长3分钟约4分钟约4分钟片段级编辑支持（替换/延展/改写）有限有限AI音乐视频内置（Veo/Gemini Omni）不支持不支持多语言人声支持（含日语、中文等）支持支持训练数据合规声称使用合法授权素材面临版权诉讼面临版权诉讼社交/社区功能内置发布/关注/播放列表内置社区内置社区移动端应用iOS已上线，Android即将上线有限有限价格策略基础免费，无需信用卡免费+付费订阅免费+付费订阅Flow Music的应用场景专辑前期探索：艺术家在策划新专辑时快速实验不同流派方向和编曲概念。影视与播客配乐：为独立电影、短视频、播客节目生成定制化背景音乐和主题歌曲。广告音频定制：根据品牌调性生成多版本广告配乐，通过AI Covers快速输出不同风格备选。音乐教育辅助：学习者通过AI理解不同曲风结构、编曲技巧和声乐处理方式。互动音乐开发：开发者基于Spaces创建音乐游戏、DJ打击垫、虚拟乐器等交互式音频体验。Remix与翻唱创作：将现有作品转换为Lo-Fi、古典、电子等风格版本，用于二次创作和采样。©️版权声明：若无特殊声明，本站所有文章版权均归AI工具集原创和所有，未经许可，任何个人、媒体、网站、团体不得转载、抄袭或以其他方式复制发表本站内容，或在非我站所属的服务器上建立镜像。否则，我站将依法保留追究相关法律责任的权利。类似于Flow Music的工具Speechify超5000万人都在用的文字转语音朗读器TextToSpeech完全免费的AI文字转语音工具天工SkyMusic昆仑万维发布的国内首个AI音乐生成大模型TurboScribe专业 AI 音视频转文字工具音虫内置AI音乐编曲的音乐制作工具音剪喜马拉雅推出的一站式AI音频创作平台暂无评论再想想发表评论暂无评论…