AI NEWSLatest AI NewsArticleMeta’s Strongest Model Muse Spark 1.3 Released: Coding Capabilities Outperform GPT-5.6 Sol, Price Remains UnchangedPublished in Latest AI NewsTime :Sep 3, 2026Read :4minuteMeta Platforms released its strongest AI model, Muse Spark 1.3, on September 2, 2026, specifically designed to extend agent workflows and enhance coding capabilities. Alexandr Wang, Meta’s Chief AI Officer, stated that the model surpasses OpenAI’s GPT-5.6 Sol in coding ability and matches the performance of Anthropic’s Claude Fable 5.1.Cleaner Code, Fewer Tokens, Long-Term Agent Tasks as the Main BattlefieldCompared to Muse Spark 1.2, version 1.3 reduces unnecessary code iterations, making the code cleaner and the overall coding style more clear: tool calls are reduced by about 20%, and the number of tokens needed to complete the same task is reduced by 25%.Muse Spark 1.3 is optimized for long-term agent tasks, enabling collaboration on multiple workflows within a single long thread, proactively correcting plan gaps, and tracking learning outcomes. When faced with ambiguous instructions, it actively asks for clarification; when encountering obstacles, it requests user assistance; and before performing critical operations, it seeks confirmation first.In addition, the reliability of the new model in handling complex long instructions has significantly improved, retaining detailed requirements and not losing constraints in multi-step tasks; multi-task processing capabilities have also been enhanced, allowing more accurate mapping of new instructions to the correct task within a single thread. The model has a stronger understanding of its own abilities and limitations, thereby reducing hallucination outputs.Pricing Remains the Same as the Previous Version, Will Be Gradually Integrated into Instagram, FacebookIn terms of pricing, Muse Spark 1.3 maintains the same price as its predecessor, Muse Spark 1.2, and offers paid access to developers through the Meta Model API: $1.25 (approximately 8.4 RMB) per 1 million tokens input, $0.15 (approximately 1 RMB) per 1 million tokens cache hit input, and $4.25 (approximately 28.6 RMB) per 1 million tokens output.Wang revealed that the adoption of the API platform is strong, with some developers using up to trillions of tokens per week. The new model will be gradually integrated into Instagram, Facebook, and the Meta AI assistant.

AI NEWSLatest AI NewsArticleOfficial Launch of the Shuiyue Yuyue Dark Side Flat Magnetic Flagship Headphones, Price: 19999 YuanPublished in Latest AI NewsTime :Sep 3, 2026Read :5minuteThe high-end audio market has welcomed a highly anticipated new product. The “Moon’s Dark Side,” the flagship planar magnetic over-ear headphones from Shuiyueyu, launched on September 2nd, with an official price of 19,999 yuan. This headphone achieved multiple breakthroughs in acoustic structure and materials, attracting the attention of many audiophiles.As the pinnacle of Shuiyueyu’s exploration of ultimate sound performance, the “Moon’s Dark Side” features a 100mm ultra-large planar magnetic driver at its core. In terms of diaphragm technology, this headphone uses an ultra-thin diaphragm with a thickness of only 500nm, significantly reducing the mass of the vibration system while maintaining excellent rigidity, effectively reducing the interference caused by the diaphragm’s own inertia on sound playback, and delivering exceptional transient response speed.To further ensure sound stability under large dynamics, the diaphragm’s conductive part innovatively uses a flexible pure gold etched circuit and introduces a secondary stress balance technology for the diaphragm, significantly improving the uniformity of stress on the diaphragm surface. Meanwhile, this headphone is equipped with Shuiyueyu’s exclusive FDT full-range drive technology, precisely placing the effective vibration area of the entire diaphragm in a uniform plane magnetic field, making the effective circuit drive area as high as 95.5%. This design fundamentally reduces segmentation distortion caused by different areas vibrating out of sync, greatly enhancing high-frequency extension and overall sound purity.In terms of body structure and craftsmanship, the “Moon’s Dark Side” also demonstrates a level of craftsmanship worthy of a flagship product. Its headphone housing and driver use a PMMA acrylic material in an integrated structure, manufactured through CNC precision machining and pure manual polishing, not only possessing anti-aging capability and excellent rigidity far exceeding ordinary plastics, but also keeping the weight within 46% of aluminum alloy density. The external frame of the headphones uses a multi-layer composite carbon fiber structure, combined with 3D-printed adjustable head pads and soft and comfortable lamb leather ear pads, perfectly balancing structural strength, durability, and comfort for long-term wearing experience.In terms of playability and cabling support, the product adopts a dual 3.5mm replaceable cable interface with an incline angle, convenient for users to freely match different front-end devices. The package includes a 4.4mm balanced headphone cable with a main core made of silver, platinum, and palladium alloy, supplemented by three-layer coaxial weaving and a double-layer shielding structure composed of high-purity copper and copper-silver plated wires, effectively reducing external electromagnetic interference. Additionally, it comes with a 6.35mm adapter cable and an XLR four-core balanced adapter cable made of the same material, fully compatible with various high-end Hi-End playback equipment, eliminating the need for users to purchase additional adapter accessories separately.

AI NEWSLatest AI NewsArticleMinistry of Justice Makes First Public Statement on AI Copyright: Supports OpenAI’s View that AI Training Constitutes Fair UsePublished in Latest AI NewsTime :Sep 3, 2026Read :6minuteThe U.S. Department of Justice (DOJ) submitted a “Statement of Interest” to the U.S. District Court for the Southern District of New York on September 1 local time, formally intervening in the lawsuit filed by The New York Times against OpenAI over copyright infringement. The DOJ clearly supported OpenAI’s position, arguing that AI companies using copyrighted material to train large language models constitutes “fair use” and does not constitute copyright infringement. The document was signed by officials including Deputy Attorney General Stanley Woodward Jr.Supporting OpenAI on the grounds of “National Security” and “Transformative Use"The DOJ first argued the necessity of supporting OpenAI from the perspective of national security, stating that if copyright rules significantly increase the difficulty of developing large language models in the United States, it could harm the competitiveness of the U.S. AI industry and pose national security risks. The document stated, “Legal rules that make establishing a strong AI industry in the United States significantly more difficult therefore threaten national security and provide a competitive advantage to foreign rivals not subject to such constraints.” The Department of Justice also cited an executive order by Trump regarding “removing obstacles to the United States’ leadership in AI,” stating that AI leadership is crucial for “promoting national security, prosperity, and economic mobility for all Americans.“The DOJ further explained that the use of copyrighted works for training large language models has a “transformative” purpose, meeting the criteria for fair use, and that “the creativity and public benefit brought by large language model training far exceed any competitive harm.” The document specifically noted that if the court supports The New York Times’ argument, it would “hinder creativity and scientific progress while hindering America’s prosperity and economic mobility.” Woodward called this document a “historically significant statement of interest.“Plaintiff’s Strong Reaction, Judge Requests Motions by September 4Graham James, a spokesperson for The New York Times, criticized the DOJ’s stance: “The government is standing with a handful of trillion-dollar AI companies while sacrificing the rights of countless American creators whose works have been stolen.” He stated that AI and creators can both thrive—AI companies just need to fairly pay for the content that made their products possible, as required by copyright law.Steven Lieberman, chief lawyer for The New York Daily News, pointed out that the government’s position “ignores the copyright provisions in the U.S. Constitution” and contradicts the views previously expressed by the U.S. Copyright Office.The New York Times filed a lawsuit against OpenAI and its partner Microsoft in December 2023, accusing the two companies of illegally using millions of copyrighted news articles to train AI systems and seeking “billions of dollars” in damages. Since then, several other news organizations, including The New York Daily News and the Chicago Tribune, along with numerous best-selling authors, have joined this consolidated lawsuit. OpenAI and Microsoft denied the allegations, asserting that their use falls within the scope of fair use.The DOJ’s intervention has escalated this landmark copyright case, and the outcome of the case may establish an important precedent for the legality of AI companies using copyrighted materials to train models. The judge presiding over the case, Sidney Stein, requested that both parties submit motions for summary judgment no later than September 4 local time.

Join NowENLatest AI NewsAI Computing Power Boom Gives Rise to Super Large Orders: Nscale Secures a $45 Billion Order from Anthropic, Contract Revenue Surpasses the $100 Billion MarkAccording to The Information, after UK cloud provider Nscale secured a $45 billion compute deal with AI firm Anthropic, its disclosed total contract revenue rose from about $51 billion to about $103 billion, surpassing $100 billion for the first time…..just now6.2KMinistry of Justice Makes First Public Statement on AI Copyright: Supports OpenAI’s View that AI Training Constitutes Fair UseThe U.S. Justice Department filed a statement of interest in Manhattan federal court, backing OpenAI in NYT v. OpenAI. It argues training LLMs on copyrighted material is fair use, citing national security and transformative use. Signed by Deputy AG Woodward and others…..just now7.4KOfficial Launch of the Shuiyue Yuyue Dark Side Flat Magnetic Flagship Headphones, Price: 19999 YuanMoondrop’s flagship planar magnetic headphones “Dark Side of the Moon” launch Sept 2 at 19,999 yuan, targeting Hi-End. Features 100mm planar magnetic driver and 500nm ultra-thin diaphragm for ultimate sound, aimed at high-end audiophiles…..just now6.4KMeta’s Strongest Model Muse Spark 1.3 Released: Coding Capabilities Outperform GPT-5.6 Sol, Price Remains UnchangedMeta released its strongest AI model, Muse Spark 1.3, focusing on long-term intelligent agent workflows and coding capabilities. Its coding performance surpasses OpenAI’s GPT-5.6 Sol and is on par with Claude Fable 5.1; compared to the previous version, it features more concise code, fewer tokens consumed, less redundant iterations, and clearer tool calls.just now8.2KMoon’s Dark Side Secretly Submits A1 Document to HKEx, Initiates IPO and Seeks Funding with a Valuation of $50 BillionAI unicorn Moonshot AI (Kimi) has confidentially filed A1 documents with HKEX to launch a Hong Kong IPO, while advancing a final pre-IPO round at a $50B pre-money valuation; the company declined to comment. It released K2.5, K2.6, and K3 this year, iterating about every three months, with ARR rising rapidly since early March…..just now4.8KValuation to Surge to $50 Billion: Moonshot AI Secretly Filed for Listing on Hong Kong Stock Exchange, AI Unicorn Accelerates Embrace of Capital MarketMoonshot AI has been reported to have submitted a confidential A1 listing application to the Hong Kong Stock Exchange, fully launching its HK IPO. The company has not commented on the report. The A1 is the official filing document for main board listing, and the company is simultaneously advancing other capital operations.just now6.5KAI Detects Early Signs of Solar Storms 9.24 Hours in Advance: New Model Captures Early Warning Signals from Acoustic Waves and Magnetic Field ChangesA team from New Jersey Institute of Technology developed the machine learning model EarlyDetect, which uses Transformer to analyze solar acoustic activity and magnetic field change data (SDO) to identify early signs of solar active regions. Although the emergence of active regions takes only a few hours, their full development can take one to several days. This model has the potential to provide earlier supplementary information for space weather forecasting.just now5.7KGoogle Pics Launches as Google’s Generative AI Image Tool to Compete with Canva and Adobe ExpressGoogle launched Google Pics, a generative image tool in Google Workspace for enterprise and AI Pro/Ultra users, powered by its Nano Banana image generation/editing model. Unlike Canva’s template ecosystem or Adobe Express’s manual pixel design, it centers on AI generation/editing for efficient creative design…..just now7.1KDoubao Update: Parallel Agents Improve Efficiency Mac Version Allows AI to Take Over the DesktopDoubao Work updated on Sept 2, launching multi-agent parallel capability for batch tasks, modular generation, and multi-dimensional research. Computer operation now extends to Mac. These upgrades evolve the AI assistant from a chatbot into an autonomous agent with desktop control…..just now6.2KU.S. Government Supports OpenAI: AI Training is a Fair Use, Limiting Copyright Would Harm Technological InnovationThe U.S. government submitted documents to the court, publicly supporting OpenAI’s position in its copyright dispute with publishers such as The New York Times. The Department of Justice stated that using news content to train AI models complies with the ‘fair use’ principle. Previously, The New York Times sued Microsoft and OpenAI in December 2023, accusing ChatGPT and Copilot of using its content without authorization.just now4.7KRelease of Gemini 3.8 Flash, Six Weeks of Continuous Updates but Ineffective WorkGoogle launched three Flash models within six weeks, with the latest Gemini 3.8 Flash targeting long-horizon software engineering, autonomous agents, and complex enterprise workflows, filling the Pro gap. Official benchmarks show strong results, especially on DeepSWE v1.1…..just n

Join NowENPremium Membership ·Limited-Time Pricing— Save TodayChoose a monthly plan that fits your needs. Each tier includes tailored points allowances and flexible usage limits for confident monitoring.Your current plan isFreeEssentialStart with essential GEO monitoring and analysis$14.9/30 daysGet PremiumPoints IncludedOne-time points 7,000Daily bonus points 100AI Query Mining50 usesGEO Rank Check50 uses5 AI platforms:GEO Brand Score Audit3 uses5 AI platforms:GEO Link Citation Tracking5 uses5 AI platforms:GEO Rank TrackingNot includedData Export：Not includedReport Generation：Not included7-Day Featured Slot in AI DirectoryNot includedDedicated SupportNot includedMost PopularGrowthBuilt for growing brands and daily GEO operations$45/30 daysGet PremiumPoints IncludedOne-time points 25,000Daily bonus points 200AI Query Mining150 usesGEO Rank Check100 uses7 AI platforms:GEO Brand Score Audit8 uses7 AI platforms:GEO Link Citation Tracking15 uses7 AI platforms:GEO Rank TrackingNot includedData Export：Not includedReport Generation：Not included7-Day Featured Slot in AI DirectoryNot includedDedicated SupportNot includedProProfessional GEO monitoring and continuous growth analysis$109/30 daysGet PremiumPoints IncludedOne-time points 65,000Daily bonus points 300AI Query Mining300 usesGEO Rank Check300 uses10 AI platforms:GEO Brand Score Audit20 uses10 AI platforms:GEO Link Citation Tracking30 uses10 AI platforms:GEO Rank TrackingUnlimited10 AI platforms:Data Export：IncludedReport Generation：Included7-Day Featured Slot in AI DirectoryNot includedDedicated SupportNot includedEnterprise PlanEnterpriseScalable GEO operations for enterprises and agencies$199/30 daysGet PremiumPoints IncludedOne-time points 120,000Daily bonus points 500AI Query Mining800 usesGEO Rank Check600 uses10 AI platforms:GEO Brand Score Audit30 uses10 AI platforms:GEO Link Citation Tracking60 uses10 AI platforms:GEO Rank TrackingUnlimited10 AI platforms:Data Export：IncludedReport Generation：Included7-Day Featured Slot in AI Directory1 usesDedicated SupportIncludedCompare PlansCompare usage allowances and premium benefits across plans at a glanceFree$0Get PremiumEssential$99/30 daysGet PremiumGrowth$299/30 daysGet PremiumPro$699/30 daysGet PremiumEnterprise$1,299/30 daysGet PremiumPoints IncludedOne-time points07,00025,00065,000120,000Daily bonus points100100200300500AI Query MiningQuery allowance550150300800GEO Rank CheckQuery allowance350100300600AI platformsGEO Brand Score AuditQuery allowance1382030AI platformsGEO Link Citation TrackingQuery allowance15153060AI platformsGEO Rank TrackingGEO Rank TrackingNot includedNot includedNot includedUnlimitedUnlimitedData ExportNot includedNot includedNot includedIncludedIncludedReport GenerationNot includedNot includedNot includedIncludedIncluded7-Day Featured Slot in AI Directory7-Day Featured Slot in AI DirectoryNot includedNot includedNot includedNot included1Dedicated SupportDedicated SupportNot includedNot includedNot includedNot includedIncludedFrequently Asked QuestionsCommon questions about Aibase memberships, credits, GEO monitoring, and billingWhat is the difference between monthly membership credits and daily login bonus credits?Monthly membership credits are the core AI computing allowance included with your membership and are issued once per membership benefit cycle.Daily login bonus credits are an additional benefit. They are automatically issued when a member first signs in to or visits Aibase on a given day, with no manual claim required.Daily bonus credits are valid only for the day they are issued and do not roll over to the next day.Do I need to manually claim daily bonus credits? Are missed days credited later?No manual claim is required. Daily bonus credits are automatically issued when you first sign in to or visit Aibase that day.If you do not sign in to or visit Aibase on a given day, that day’s bonus credits will not be issued and will not be credited retroactively.Daily bonus credits are an additional membership benefit and do not affect your regular monthly membership credits.Do monthly membership credits roll over to the next month?No. Membership credits are issued in 30-day benefit cycles.Unused membership credits expire at the end of the current benefit cycle and do not roll over into the next cycle.At the beginning of a new benefit cycle, the credit allowance for your current membership plan is issued again.If I buy a 3-, 6-, or 9-month membership, will all credits and usage quotas be issued at once?No. A multi-month purchase provides membership access for the selected duration; it does not issue all future credits and usage quotas upfront.Membership benefits are still managed in 30-day cycles. Credits are issued and tool usage quotas are reset at the start of each cycle.For example, a 3-month membership consists of three consecutive 30-day benefit cycles.In what order are different credit balances used?Credits are deducted in the following order: daily login bonus credits →

Muse Voice Transcribe – Meta 推出的首个实时音频感知模型AI工具23小时前发布AI小集02Muse Voice Transcribe是什么Muse Voice Transcribe 是 Meta 推出的首个实时音频感知模型，集流式自动语音识别、说话人分割与端点检测于一体。模型支持 70 多种语言及原生语码切换，可处理超 1 小时长音频与 20 人以上同时对话，在流式转写和语音分割基准测试中均排名第一。Muse Voice Transcribe 基于强化学习的自适应延迟技术，能动态权衡速度与准确率。Muse Voice Transcribe 已集成至 Meta AI 与 Muse Code，定位为 AI 系统的实时感知层。Muse Voice Transcribe的主要功能实时流式 ASR：模型以 80ms 为单位实时处理音频并输出文字，无需等待整段音频结束。说话人分割：自动识别并标记超过 20 位不同说话人，区分谁在何时发言。语音端点检测：精准判断语音开始与结束时刻，支持自然停顿与连续对话。多语言与语码切换：覆盖 70 多种语言，原生支持中英文夹杂等任意语码转换。长音频与多人会议：可连续处理超过 1 小时音频，无需后处理即可输出完整转写。Muse Voice Transcribe的技术原理自回归多模态架构：基于 Muse Spark 系列的自回归模型，将音频流逐帧转化为软 token 并与文本 token 统一建模，实现音频到文本的端到端生成。流式音频 token 机制：每 80ms 音频片段编码为软 token，模型通过预测<|next_audio|>决定是否继续监听下一段，收到<|empty_audio|>则立即输出剩余文本，实现真正的流式推理。任务扩展特殊 token：在 ASR 基础上引入<|start_of_turn|>与<|speaker_{A-Z}|>实现说话人切换识别，引入<|speech_onset|>与<|speech_endpoint|>实现端点检测，所有任务共享同一流式训练框架。强化学习自适应延迟：通过 RL 将词错误率奖励与延迟奖励以乘法方式组合，使模型根据每个单词的识别难度动态决定等待更多上下文还是立即输出，在速度与准确率间达到帕累托最优。如何使用Muse Voice Transcribe选择入口：通过 Meta Model API、Mac 版Meta AI或Muse Code三种方式接入服务。启动语音：在 Muse Code 中按下⌘ + v快捷键，或在 Mac 版 Meta AI 中点击语音按钮开始听写。跨应用使用：在任意 Mac 应用中按住 Fn 键，即可调用 Meta AI 完成语音转写。API 调用：通过 Meta Model API 按量计费（每 1000 分钟 3 美元），集成到自己的应用或工作流中。Muse Voice Transcribe的核心优势实时流式 ASR：用 80ms 为单位处理音频，实现真正的低延迟实时转写。多人说话人分割：原生支持超过 20 位说话人区分，无需后处理即可精准归属每句话。自适应延迟：通过强化学习动态调整每个单词的等待时间，在速度与准确率之间达到帕累托最优。无缝语码切换：原生支持句子内或句子间的任意语言切换，精准识别中英文夹杂等混合表达。超长音频处理：支持超过 1 小时的连续长音频输入，适用于长时间会议或访谈场景。语音端点检测：自动标记语音开始（<|speech_onset|>）与结束（<|speech_endpoint|>），精准判断用户何时说完。基准测试第一：在 Artificial Analysis 流式语音转文本和公开语音分割基准测试中均排名第一。Muse Voice Transcribe的项目地址项目官网：https://research.meta.ai/blog/introducing-muse-voice-transcribeMuse Voice Transcribe的同类竞品对比对比维度Muse Voice Transcribe(Meta)Gemini 3.1 Flash Live(Google)产品定位专门的实时音频感知与 ASR 模型通用实时多模态对话模型，ASR 只是其输入能力之一核心能力流式 ASR、说话人分割、语音端点检测实时双向语音对话（ASR + TTS）、多模态（音频/视频/图像/文本）、函数调用词错误率AA-WER Streaming Index3.1%，Artificial Analysis 排名第一未公布具体 ASR WER 数值，主打低延迟流式 ASR 与对话流畅度说话人分割原生支持20+ 位说话人，DER 仅17.5%未明确支持多人说话人分割，专注单轮/双轮实时对话延迟策略自适应延迟，通过强化学习为每个单词动态调整等待时间Sub-200ms固定首 token 延迟，追求极速响应长音频处理原生支持1 小时+连续音频，无需分段纯音频会话默认15 分钟限制，需接入 session 续接机制语码切换原生支持句子内/间任意语码转换，精准识别中英文夹杂支持 90+ 种语言，但未强调语码切换专项优化纯转写成本$0.18/小时音频输入$0.30/小时+ 文本输出约$0.07/小时，合计约$0.37/小时输出能力仅文本转写 + 说话人标签（<|speaker_A|>）文本 + 音频合成（TTS）+ 工具调用 + 结构化输出 + 实时搜索Muse Voice Transcribe的应用场景跨国会议实时纪要：支持20+人同时发言、超过1小时的长时间会议，自动区分说话人并精准转写中英文夹杂的讨论内容。AI语音助手/Agent：作为实时感知层，提供低延迟流式ASR与精准的语音端点检测，让AI助手知道用户何时开始和结束说话。播客与访谈字幕生成：处理长达数小时的音频素材，原生支持多人对话分割，无需后期手动标注说话人。多语言客服质检：实时转写客服与客户的混合语言对话，用上下文偏好提升专业术语识别准确率，并自动归档。开发者语音输入工具：通过Mac版Meta AI或Muse Code的Fn键全局调用，在任何应用中实现语音转写编程、写文档或发消息。# 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Atlas - World Labs 推出的全球首个多模态世界模型下一篇Hy4 preview 轻量版 - 腾讯混元开源的Hy4量化压缩模型相关文章TinyR1-Preview – 奇虎360联合北大团队推出的推理模型AI小集2ebook2audiobookXTTS – 开源电子书转有声书 AI 工具，支持16种语言AI小集2Stable Virtual Camera – Stability AI 等机构推出的 AI 模型，2D图像转3D视频AI小集2OmniAudio-2.6B – Nexa AI推出的端侧多模态音频语言模型AI小集3POINTS 1.5 – 腾讯微信推出的多模态大模型AI小集3Huxe AI – AI音频伴侣应用，提供个性化音频服务AI小集4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Lovart美图设计室绘蛙AIupdream办公小浣熊秒哒最新收录ByteCPAgent.SpaceQwenWorkDeepSkillAlphaVowXCell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！45分钟前Easel – 浙大联合北大开源的 AI 社交媒体内容工作台51分钟前Gemini 3.8 – 谷歌推出的推理与编码模型56分钟前Fusion Model – FUMO Lab 推出的智能分配模型16小时前Hy4 preview 轻量版 – 腾讯混元开源的Hy4量化压缩模型21小时前Atlas – World Labs 推出的全球首个多模态世界模型23小时前Claude Mythos 5.1 – Anthropic 推出的 Claude 系列旗舰模型23小时前Claude Fable 5.1 – Anthropic 推出的最新旗舰大模型23小时前WeMM-Embedding – 腾讯开源通用多模态Embedding模型2天前FireRedAudio – 小红书 FireRed 推出的通用音频语言模型2天前MiniMax H3 Max – MiniMax 推出的实时视频生成模型3天前Ling-3.0-flash-Fin – 蚂蚁百灵推出的首个金融增强模型3天前AI 技能成千上万怎么选？DeepSkill 精准搜索实战3天前Rome – 开源的智能体操作系统，自然语言生成完整应用4天前Headlong – Laude Institute 开源的 AI Agent 微框架4天前

IT之家9 月 2 日消息，科技媒体 9to5Mac 昨日（9 月 1 日）发布博文，报道称 Meta 公司推出 Muse Voice Transcribe，这是其首个实时音频感知模型。开发者可通过 Meta Model API 调用该能力，定价为每 1000 分钟音频 3 美元（现汇率约合 20.2 元人民币），等同每小时 0.18 美元（现汇率约合 1.2 元人民币）。该模型在同一流程中整合流式自动语音识别、说话人分离与端点检测，用户说话时，系统可同步输出文字，无需等待完整录音结束后再单独处理。IT之家注：“流式”就是边说边转写，模型不必等整段音频说完才出结果，而是会一小段一小段持续输出文字，所以延迟更低，适合实时听写、会议记录、通话字幕。Muse Voice Transcribe 可在包含 20 余名说话者的录音中分离不同发言者，还能识别说话是否结束，从而自动确定一段输入的边界。模型训练覆盖 70 余种语言，其中 25 种语言在发布时完成验证。该模型支持长度超过 1 小时的音频，也支持句内或句间的自然语言切换。开发者还可通过语言、关键词和上下文偏置，提高特定语言、术语或场景下的识别效果。在技术方面，Meta 为兼顾实时响应和识别准确度，引入名为自适应延迟（adaptive delay）的动态决策机制。系统不会为所有词语采用固定的速度与准确度取舍，而是针对每个词判断需要额外接收多少音频后再输出识别结果。对于较容易识别的语音，系统可更快提交转写结果。对于发音不清、术语较多或语境复杂的词语，系统会调用更多前后文音频信息。这种机制旨在兼顾。Meta 称，截至 2026 年 9 月 1 日，Muse Voice Transcribe 位列 Artificial Analysis 流式语音转文本排行榜第 1 名。

Qwen 3.8-Max – 阿里云 Qwen 团队推出的旗舰大模型AI工具12小时前更新AI小集02Qwen 3.8-Max是什么Qwen 3.8-Max 是阿里云 Qwen 团队推出的旗舰大模型，总参数量达 2.4 万亿，基于Qwen 3.5架构扩展而成。作为 Qwen 家族迄今最强大的模型，是首个即将开源权重的 Max 级别模型。模型在编程、办公、科研及长程任务上实现全面升级，支持通过千问AI平台 API 调用，兼容 OpenAI 与 Anthropic 协议，可无缝集成至Claude Code、Codex、Qoder等主流智能体框架。Qwen团队已开源总参数2.4T、每Token激活95B的MoE大模型Qwen3.8-2.4T-A95B，原生支持256K上下文，在编程与Agent任务上表现领先，为Qwen-Max级别首次开放权重。Qwen3.8-Max-0902 版本通过针对编程与专业办公场景的强化后训练，在 CodeArena 前端编程总榜夺冠，同时以每百万 Tokens 仅 5 美元的性价比刷新全球模型新上限。Qwen 3.8-Max的主要功能长程自动编程：支持十天级自主编程，构建自进化 harness 并完成需求收集、issue 派发、代码生成与自我修复。论文复现与超越：独立复现论文实验后，通过四轮自我进化循环测试 18 种改进想法，在 AIME24 上超越原方法 2.7 分。竞赛实战：24 小时内独立搭建并迭代方案，准确率从 0.60 升至 0.853，击败 87% 的人类参赛队伍。办公提效：覆盖数百种高价值职业场景，如律师一小时标出 1,284 条条款、设计师一次成型 8 页可交互原型。量化策略研发：从单句描述出发调度约 330 个子智能体，完成约 6,000 次回测，交付端到端 ETF 轮动策略。芯片自主设计：历经约 500 轮交互，将密码加速器网表门数从 8,298 优化至 678，面积缩减 81% 并实现时序闭合。Qwen 3.8-Max的技术原理架构扩展：基于 Qwen 3.5 架构基础扩展而成，总参数量达 2.4 万亿，激活参数 95B，在保持架构继承性的同时实现规模跃升。真实世界强化学习系统：通过联合扩展 RL 环境数量与训练算力提升通用工作能力，核心解决三个耦合挑战：在任务、工作空间、Harness三个维度上解耦合，使环境数以组合方式自然增长；构建统一奖励系统，将基于执行的校验、文本及渲染后视觉输出的 rubric 评估、agentic 检查统一内化，消除任务专用验证器的不一致性；构建在线数据均衡器，对每个 batch 在任务、难度、工作区与 harness 上重构均衡分布，降低梯度方差，支撑训练算力稳定持续扩展。自进化反馈闭环：模型通过执行-反馈-迭代的自适应机制推进优化，在数百至数千轮交互中保持高度连贯的系统性策略，依赖仿真、综合、布局等反馈闭环，实现算法级数据通路重构而非浅层语法微调。动态工作流编排：用编程方式驱动任务规划，将编排逻辑固化为可复现程序，支持从单链路连贯研发到大规模并行探索的灵活切换，使复杂任务的串行推进转化为一次对话内可规模化交付的自动化闭环。如何使用Qwen 3.8-Max千问AI平台调用：注册账号获取 API Key，通过兼容 OpenAI或 Anthropic 协议的接口直接调用模型。调节推理深度：通过reasoning_effort参数设置xhigh（默认，深度分析）、medium（平衡）或low（高效推理），按需控制成本与性能。集成智能体框架：配置 API Key 与 Base URL 后，可接入 Claude Code、Codex、Qoder CLI、Qwen Code、OpenClaw 等主流编程助手与 Agent 框架。开源权重部署：下周起从 Hugging Face 或 ModelScope 下载模型权重，进行本地部署与二次开发。Qwen 3.8-Max 的核心优势超长程自主执行：可连续专注数天至数十天，通过执行-反馈-迭代闭环自我进化，全程无需人工介入。真实世界 RL 训练：联合扩展环境数量与训练算力，在 QwenWork、Claude Code、Codex、OpenClaw、Hermes 等 harness 上表现均衡且持续提升。端到端交付能力：通过动态工作流将编排逻辑固化为可复现程序，把原本需数周串行推进的复杂任务压缩为一次对话内可规模化交付的自动化闭环。Qwen 3.8-Max 的项目地址项目官网：https://qwen.ai/blog?id=qwen3.8魔搭模型库：https://modelscope.cn/models/Qwen/Qwen3.8-2.4T-A95BQwen 3.8-Max 的同类竞品对比维度Qwen 3.8-MaxGPT 5.6 Sol论文复现PaperBench 93.0 分，独立复现后自我进化超越原文PaperBench 90.5 分软件工程SWE-bench Pro 67.7 分，FrontierSWE 73.5 分SWE-bench Pro 64.6 分，FrontierSWE 无官方数据终端编码Terminal Bench 2.1 86.6 分Terminal Bench 2.1 88.8 分通用办公CoWorkBench 74.8 分，JobBench 53.4 分CoWorkBench 71.5 分，JobBench 45.4 分多模态推理MMMU-Pro 82.3 分，BabyVision 82.0 分MMMU-Pro 83.0 分，BabyVision 65.5 分视觉智能体OSWorld-Verified 86.1 分，AndroidWorld 85.3 分OSWorld-Verified 83.2 分，AndroidWorld 77.6 分视频理解VideoMME 90.4 分，MLVU 90.8 分VideoMME 89.5 分，MLVU 87.6 分长程任务365 天电商模拟 4.16 倍回报，芯片设计 500 轮优化无公开同类长程自主任务数据Qwen 3.8-Max 的应用场景法律合规审查：单次通读数百份文档，一小时内标出上千条相关条款，替代法务团队约一周工作量。UI/UX 原型设计：为数字银行 App 一次生成 8 页高保真可交互原型，无需人工返修。餐饮菜单开发：基于上百份食材资料一次性产出 26 道菜的完整菜单，控制食材成本率在 33.8%。结构工程建模：仅凭一份图纸在浏览器中还原 30 层写字楼抗震结构模型，支持实时悬停查看关键指标。康复医学可视化：将 2D 纸质评估单转化为可自由旋转、逐层浮现的 3D 交互演示，帮助患者理解康复路径。# 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MindMemOS - 华为诺亚方舟开源的 AI Agent 记忆操作系统下一篇GenOffice - Genspark 推出的本地 AI Office 客户端相关文章Saner.AI – AI知识管理工具，提效的个人AI笔记和搜索引擎AI小集2Creao.ai – AI应用开发平台，自主生成、优化和部署工作流AI小集3Symphony – OpenAI 开源的 Agent 编排系统AI小集3Kimi公文笔杆子 – Kimi推出的AI公文写作助手AI小集5PaddleSpeech – 百度飞桨团队开源的语音处理工具AI小集3LinGen – Meta联合普林斯顿大学推出的文本到视频生成框架AI小集2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Lovart美图设计室绘蛙AIupdream办公小浣熊秒哒最新收录ByteCPAgent.SpaceQwenWorkDeepSkillAlphaVowXCell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！56分钟前Easel – 浙大联合北大开源的 AI 社交媒体内容工作台1小时前Gemini 3.8 – 谷歌推出的推理与编码模型1小时前Fusion Model – FUMO Lab 推出的智能分配模型17小时前Hy4 preview 轻量版 – 腾讯混元开源的Hy4量化压缩模型21小时前Muse Voice Transcribe – Meta 推出的首个实时音频感知模型23小时前Atlas – World Labs 推出的全球首个多模态世界模型23小时前Claude Mythos 5.1 – Anthropic 推出的 Claude 系列旗舰模型23小时前Claude Fable 5.1 – Anthropic 推出的最新旗舰大模型23小时前WeMM-Embedding – 腾讯开源通用多模态Embedding模型2天前FireRedAudio – 小红书 FireRed 推出的通用音频语言模型2天前MiniMax H3 Max – MiniMax 推出的实时视频生成模型3天前Ling-3.0-flash-Fin – 蚂蚁百灵推出的首个金融增强模型3天前AI 技能成千上万怎么选？DeepSkill 精准搜索实战3天前Rome – 开源的智能体操作系统，自然语言生成完整应用4天前

Kimi K2.7 Code – 月之暗面开源的新一代编程专用模型AI工具3个月前更新AI小集03Kimi K2.7 Code是什么Kimi K2.7 Code 是月之暗面（Moonshot AI）开源的新一代编程专用模型。相比K2.6，在长上下文编程场景的指令遵循能力、长程任务性能上显著提升，改善了过度思考倾向，平均 token 消耗减少 30%。在 Kimi Code Bench v2、Program-Bench、MLS Bench Lite 等基准测试中性能分别提升 21.8%、11% 和 31.5%，Agent 自主执行能力也提升约 10%。Kimi K2.7 Code的主要功能长上下文编程能力跃升：显著提升长上下文编程场景的指令遵循能力，长程编程任务性能表现大幅增强。Token 效率优化：改善长程任务中的过度思考倾向，平均 token 消耗减少30%，以更少 token 达到更高性能。代码基准全面突破：在 Kimi Code Bench v2、Program-Bench、MLS Bench Lite 等内外部基准测试中性能显著提升（分别提升 21.8%、11%、31.5%）。Agent 自主执行能力进化：在 Kimi Claw 24/7 Bench、MCP Atlas、MCP Mark Verified 等 Agent 基准测试中性能提升约10%。思考模式（Thinking）：必须开启思考模式才能发挥最佳性能，API 和 Kimi Code 均默认开启。6 倍速高速版支持：常规编程场景输出速度约180 Token/s，短上下文可达260 Token/s，6 月 15 日上线 API。开源与本地部署：模型已开源并上架 Hugging Face，支持开发者本地部署。多平台接入：支持通过 Kimi API 开放平台、Kimi Code 工具、Kimi 会员及企业版接入使用。缓存命中优化：支持缓存机制，缓存命中输入价格降至1.3 元/1M tokens，有效降低调用成本。如何使用Kimi K2.7 Code通过 Kimi Code 工具使用：访问kimi code，默认模型已自动升级为 Kimi K2.7 Code，可直接开始编程辅助。通过 Kimi API 开放平台调用：访问 Kimi 开放平台 https://platform.kimi.com/，接入 API 创建应用，参考快速开始文档进行集成。通过会员/企业版体验：订阅 Kimi Code Plan 或企业版 Kimi 会员计划（包含 Kimi Code Plan 权益），即可使用新模型。本地部署模型：前往 Hugging Face（moonshotai）下载模型权重，自行在本地环境部署运行。开启思考模式（Thinking）：使用 K2.7 Code 时必须打开思考模式以发挥最佳性能；API 和 Kimi Code 均默认开启，若手动关闭则 API 会报错、Kimi Code 会回退到 K2.6 模型。使用高速版（6月15日起）：通过 Kimi API 开放平台调用高速版模型，享受约5-6 倍输出速度（常规场景180 Token/s，短上下文可达260 Token/s）。微信关注回复“开源”，加入AI开源项目交流群Kimi K2.7 Code的核心优势性能全面跃升：Kimi Code Bench v2 提升21.8%，Program-Bench 提升11%，MLS Bench Lite 提升31.5%，代码能力显著超越前代。Token 效率革命：长程任务过度思考倾向大幅改善，平均 token 消耗减少30%，以更少成本获得更高输出质量。Agent 自主执行进化：Kimi Claw 24/7 Bench、MCP Atlas 等 Agent 基准测试性能提升约10%，自主化任务执行能力更强。6 倍速极速体验：高速版常规编程场景输出速度达180 Token/s，短上下文峰值可达260 Token/s，仅需 2 倍价格。极致性价比：1M token 标准输入6.5 元、输出27 元，与 K2.6 保持一致；缓存命中输入仅1.3 元，配合限时充赠最高返30%。Kimi K2.7 Code的同类竞品对比对比维度Kimi K2.7 CodeKimi K2.6GPT-5.5(xhigh)Opus 4.8(xhigh)Kimi Code Bench v262.050.969.067.4Program Bench53.648.369.163.8MLS Bench Lite35.126.735.542.8Kimi Claw 24/7 Bench46.942.952.850.4MCP Atlas76.069.479.481.3MCP Mark Verified81.172.892.976.4标准输入价格 (元/1M tokens)6.56.5——标准输出价格 (元/1M tokens)2727——缓存命中输入 (元/1M tokens)1.3———相对 Token 消耗比 K2.6 减少30%基准——高速版输出速度180260 Token/s———Kimi K2.7 Code的应用场景大规模代码库理解与开发：用长上下文能力处理数万行代码的复杂项目，进行跨文件分析、架构梳理与功能开发。长程软件工程任务（SWE）：端到端完成需求分析、代码编写、测试用例生成与调试修复，支持 Program-Bench 等复杂工程基准。Agent 自动化工作流：通过 MCP 协议连接外部工具与服务，实现自主化代码部署、文档生成、CI/CD 流程编排等 24/7 不间断任务。实时交互式编程辅助：借助高速版（180260 Token/s）进行快速原型开发、即时代码补全、实时错误诊断与修复建议。代码审查与重构优化：对历史代码进行深度审查，识别潜在漏洞、性能瓶颈，并执行大规模重构以降低技术债务。多模态编程场景：结合 Kimi 系列多模态能力，处理包含 UI 截图、设计稿、视频演示等视觉信息的编程任务（如前端还原、界面实现）。# 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EvoQuality - 字节跳动开源的图像质量评估模型下一篇openPangu 2.0 - 华为开源的盘古大模型升级版相关文章StockBot – 基于Llama3的AI金融Agent，提升股票投资效率AI小集3Songburst – AI音乐生成器，根据风格和情感描述生成原创音乐AI小集2Food Mood – Google AI 推出的创意菜谱生成工具AI小集2Imagine v0.9 – xAI推出的视频生成模型AI小集3Leffa – Meta 开源的图像生成框架，精确控制人物的外观和姿势AI小集2Origin – Cursor 推出的 AI 原生代码托管平台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Lovart美图设计室绘蛙AIupdream办公小浣熊秒哒最新收录ByteCPAgent.SpaceQwenWorkDeepSkillAlphaVowXCell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小时前Easel – 浙大联合北大开源的 AI 社交媒体内容工作台1小时前Gemini 3.8 – 谷歌推出的推理与编码模型1小时前Fusion Model – FUMO Lab 推出的智能分配模型17小时前Hy4 preview 轻量版 – 腾讯混元开源的Hy4量化压缩模型21小时前Muse Voice Transcribe – Meta 推出的首个实时音频感知模型23小时前Atlas – World Labs 推出的全球首个多模态世界模型23小时前Claude Mythos 5.1 – Anthropic 推出的 Claude 系列旗舰模型23小时前Claude Fable 5.1 – Anthropic 推出的最新旗舰大模型24小时前WeMM-Embedding – 腾讯开源通用多模态Embedding模型2天前FireRedAudio – 小红书 FireRed 推出的通用音频语言模型2天前MiniMax H3 Max – MiniMax 推出的实时视频生成模型3天前Ling-3.0-flash-Fin – 蚂蚁百灵推出的首个金融增强模型3天前AI 技能成千上万怎么选？DeepSkill 精准搜索实战3天前Rome – 开源的智能体操作系统，自然语言生成完整应用4天前

Claude Code是什么Claude Code 是 Anthropic 公司推出的基于命令行的AI 编程工具。Claude Code集成先进的 Claude Opus 4 语言模型，能理解自然语言指令，自动生成、修改和调试代码，并与 VS Code、JetBrains 等主流 IDE 深度集成。Claude Code支持 200K 超长上下文，能快速索引整个代码库并理解复杂依赖关系。Claude Code 具备文件操作、代码搜索、网页浏览和 Git 工作流管理等功能，显著提升开发效率。Claude Code帮助开发者快速生成高质量代码，能直接在开发环境中执行命令，避免频繁切换界面，极大地优化开发流程。最新版推出自动安全审查功能，在漏洞发布之前捕获漏洞。Claude Code 新功能 Opus 规划模式支持高效且智能的开发规划，用户能在复杂项目中借助 Opus 4.1 模型的强大分析能力制定高质量的开发计划，在执行阶段切换到 Sonnet 模型，实现高效、低成本的代码生成。Claude Code的主要功能代码库深度理解：基于智能搜索技术，快速理解整个代码库的结构和依赖关系，无需手动选择上下文文件。多文件协同修改：能够跨多个文件进行协调一致的代码修改，确保代码逻辑的完整性和一致性。代码生成与优化：专门针对代码理解和生成进行了优化，能够快速生成高质量的代码片段、测试用例和修复代码中的问题。终端集成：直接在开发者的终端中运行，无需切换工具，减少上下文切换的麻烦。与主流 IDE 集成：支持 VS Code 和 JetBrains 等主流开发工具，直接在 IDE 中提供代码建议和修改，无缝融入现有开发流程。工作流整合：与 GitHub、GitLab 等工具集成，支持从阅读问题、编写代码、运行测试到提交 PR 的全流程操作。智能编辑：基于对代码库和依赖关系的理解，进行强大的多文件编辑，确保代码修改的有效性和可靠性。用户控制与定制：不会在未经明确批准的情况下修改文件，适应开发者的编码标准和模式，并通过 SDK 或 GitHub Actions 进行定制化配置。终端运行的/security-review命令：支持用户在终端输入此命令，Claude Code将自动检查代码中的SQL注入、XSS等安全隐患，能直接修复发现的漏洞。GitHub Actions集成：在GitHub仓库配置相关文件后，每次提交新的PR时，自动触发安全审查，检测出漏洞后在代码行上添加内联注释，解释问题并给出修复建议。Opus 规划模式：在规划阶段调用强大的 Claude 4.1 Opus 模型制定高质量开发计划，在执行任务时使用高效且低成本的 Claude 4 Sonnet 模型，实现智能与效率的结合。Agent View：让开发者在一个终端界面内同时调度、监控和管理多个并行 AI 会话，无需多开窗口可实时查看各任务状态、快速切换介入或解除阻塞。Artifacts：支持将 AI 工作过程实时生成为可共享的交互式网页，方便团队同步查看与协作。如何使用Claude Code安装 Node.js：确保系统安装 Node.js 18 或更高版本。安装 CLI：在终端运行以下命令安装 Claude Code 的命令行工具：npminstall-g@anthropic/claude-cli登录和初始化：登录：运行claude login并按照提示输入 API 密钥（从 Anthropic 官网获取）。初始化：运行claude init，连接到代码库。开始使用：生成代码：用英语描述需求，例如：claude-p"Create a function to calculate the sum of two numbers in Python"修复错误：描述错误或粘贴错误消息，例如：claude-p"Fix this error: TypeError: Cannot read properties of null (reading ’length’)“自动化任务：运行命令解决繁琐任务，例如：claude-p"Resolve merge conflicts in the current branch"集成到开发环境：终端：直接在终端中使用。IDE：安装 VS Code 或 JetBrains 插件，实现在 IDE 中使用。Claude Code 的产品定价Pro 计划：年付 $17/月，月付 $20/月。Max 5x 计划：$100/人/月。Max 20x 计划：$200/人/月。Claude Code的应用场景代码理解：快速解释复杂代码的功能和逻辑，帮助开发者迅速理解项目结构。代码修改：自动修复代码中的错误或漏洞，快速定位问题并提供解决方案。代码测试：为代码自动生成单元测试或集成测试，确保代码的可靠性和稳定性。版本控制：支持自然语言完成 Git 提交、分支管理等操作，简化版本控制流程，让 Git 操作更加便捷高效。功能实现与扩展：快速实现新功能或扩展现有功能，加速开发进度，帮助开发者高效完成任务。©️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Claude Code的工具MiMo Code小米大模型团队开源的新一代 AI 编程助手JetBrains AIJetBrains推出的AI编程开发助手iFlow CLI心流AI推出的免费终端 AI 智能体Kiro亚马逊公司推出的 AI IDECodeium免费的AI编程工具，智能生成和补全代码YouWare一站式 AI 编程社区与开发平台6 条评论再想想发表评论wi游客此ide已放弃 ，国内无法使用，哎1年前 (2025)回复AI小集需要很强的魔法，还经常封号1年前 (2025)回复wi游客对。弄得很累，不如用亚马逊的ai ide，不用翻墙，也不封号1年前 (2025)回复3游客Claude 太强大了，可惜用不来6个月前回复Ai游客需要手机验证，不支持国内无866个月前回复AI小集是的，最好搞一个美国手机号6个月前回复