AI NEWSLatest AI NewsArticleNo Cloud API Required! Liquid AI Launches Local-Running LFM2.5 Edge Intelligence ModelPublished in Latest AI NewsTime :Aug 5, 2026Read :2minuteOn August 5, the AI startup Liquid AI officially announced the release of LFM2.5-2.6B, a terminal-side model designed for agent workflows. The model has 2.6 billion parameters and can run entirely locally on terminal devices such as smartphones without relying on cloud API, providing developers with a low-cost, low-latency, and privacy-protected edge-side agent solution.In terms of technical architecture and training approach, LFM2.5-2.6B is pre-trained on about 34 trillion token, with a vocabulary expanded to 128K. Its post-training process integrates four stages: supervised fine-tuning, teacher specialization, multi-domain strategy distillation, and agent reinforcement learning, giving the model excellent planning, tool calling, and complex multi-step task handling capabilities. Benchmark tests show that the model’s performance in following instructions and using tools is comparable to large models nearly four times its size.Additionally, the model performs well in hardware compatibility and inference performance, achieving a reasoning speed of 220 tokens/second on the M5Max chip, and maintaining a stable response speed of 30 tokens/second on mobile devices. Currently, LFM2.5-2.6B and its base version are fully open-sourced on Hugging Face, and natively support mainstream inference frameworks such as llama.cpp, MLX, and vLLM, marking the practical stage of high-efficiency edge-side agents.

AI NEWSLatest AI NewsArticleMeta Open-Sources 30B-Parameter Muse Glimmer for Local AI Agent on Consumer HardwarePublished in Latest AI NewsTime :Aug 11, 2026Read :3minuteMeta released the open-source model Muse Glimmer on August 10, aimed at supporting local AI agents on consumer-grade hardware, representing the latest implementation of its “Personal Superintelligence” vision. The model has 30 billion parameters and is essentially an open version of Meta’s closed model Muse Spark, released in April this year. The model weights are released under the Apache 2.0 license, allowing developers to download, modify, and fine-tune them.Muse Glimmer is designed for multi-step agent tasks, capable of calling tools, writing and debugging code, handling files and screenshots, and executing extended workflows. The model supports text and images, and its training covers over 100 languages. It can run locally on a Mac or PC with a single consumer-grade GPU, and is designed to be “always-on,” working even without an internet connection.Meta envisions its use in managing schedules, writing emails, organizing files, and other tasks that require processing large amounts of personal data. By processing information locally, it reduces the need to upload data to the cloud, exploring more privacy-focused personal AI assistants.This approach aligns with Zuckerberg’s previous “Personal Superintelligence” concept. In a new public letter released on August 10, he stated that widespread access to superintelligence could empower individuals and hopes that related tools will eventually be available for free or at an affordable price. However, Meta still differentiates between open and closed models: the more powerful Muse Spark remains closed, while Glimmer allows users to download and run it locally, showing its layered strategy between AI openness and model control.

AI NEWSLatest AI NewsArticleNVIDIA Collaborates with Six Financial Institutions to Raise Over $500 Billion for Building an AI Computing Financing PlatformPublished in Latest AI NewsTime :Aug 11, 2026Read :3minuteAugust 11th news, NVIDIA announced the establishment of an independent computing power financing platform in cooperation with six financial institutions: Apollo, BlackRock, Fortress, Brookfield, Goldman Sachs and KKR, aiming to mobilize over $50 billion in third-party capital for AI infrastructure construction. NVIDIA has signed a memorandum of understanding with these institutions on August 10th, and the cooperation will take effect after the final agreement is signed.According to the plan, NVIDIA will establish a dedicated large-scale fund pool with its partners, providing financial support at competitive financing rates to customers, helping enterprises, cloud service providers and AI development institutions obtain the computing resources needed for building AI infrastructure.This collaboration focuses on expanding AI computing infrastructure. With the continuous growth of demand for training and inference of generative AI models, data centers, GPU clusters, and energy infrastructure have become the core of industry competition, and substantial capital investment has driven technology companies and financial institutions to explore new financing models.NVIDIA hopes that this financing platform will reduce the financial pressure on customers expanding their computing power, accelerate the construction of AI “factories”, and provide more extensive computing resource support for global enterprises and developers. This move also shows that the competition in the AI industry is extending from model and chip capabilities to computing infrastructure and capital system construction. A

AI NEWSLatest AI NewsArticleTencent Opensources New Browser Automation Tool BrowserSkill: Giving AI Agents Independent Hands to WorkPublished in Latest AI NewsTime :Aug 11, 2026Read :4minuteAt a time when artificial intelligence technology is accelerating its practical application, how to make AI Agents more efficiently handle daily web tasks has always been a focus for developers. Recently, Tencent officially open-sourced a browser automation tool specifically designed for intelligent agents - BrowserSkill.According to the information, the core design of this tool is specifically aimed at various AI Agents, which is significantly different from traditional Browser Use solutions that serve as plugins for large models. In actual experience, BrowserSkill demonstrates three prominent highlights: first, it can fully reuse the real login status of the browser, so that accounts previously authorized or logged in do not need to be re-verified when the intelligent agent is called; second, the operation process does not interrupt the current work, and the Agent will execute various instructions in an independent window, completely not interfering with the user’s current web browsing or daily communication; finally, it has a high level of compatibility with mainstream agents, by providing a simple bsk command-line tool (CLI), allowing various agents such as Claude Code, Codex, and Tencent’s own Workbuddy to be seamlessly integrated.In practical scenario testing, developers conducted on-site verification through the [Workbuddy] platform. Whether it’s letting AI search for popular tech topics on Zhihu and accurately summarize core viewpoints, or scraping high-rated comments on social platforms and generating group sentiment profiles, the entire operation process can be efficiently completed within a few minutes. Due to the use of a CLI-independent drive and extended intermediary architecture at the bottom, the model does not need to directly recognize complicated web elements, thus effectively ensuring execution speed and stability.Industry insiders pointed out that compared to similar open-source solutions mainly serving human operations, the release of BrowserSkill further bridges the “last mile” of AI intelligent agents’ interaction with the real network, providing developers and content creators with a new technical path for handling multiple tasks simultaneously.Project address: https://github.com/Tencent/BrowserSkill

AI NEWSLatest AI NewsArticleClaire Expands Its Computing Power Map: Anthropic Collaborates with Riot to Achieve a $9.1 Billion Long-Term Cooperation AgreementPublished in Latest AI NewsTime :Aug 11, 2026Read :2minuteThe arms race in the field of artificial intelligence is advancing into the infrastructure of computing power. According to insiders, the AI company Anthropic has recently signed a major computing power deal worth up to $9.1 billion with Riot Platforms Inc., a cloud computing company that has transitioned from Bitcoin mining.As an important step in its business transformation, Riot Platforms Inc. has officially entered the field of artificial intelligence data center computing power services. According to the information it disclosed, the two parties have finalized a 20-year deep cooperation agreement, under which Riot Platforms Inc. will provide a total of 191 megawatts of power capacity through its modernized campus in Rockdale, Texas, to support the computing power needs of this leading AI company’s large-scale models.Industry experts point out that as competition among major AI model manufacturers for underlying computing power intensifies, transforming traditional energy and cryptocurrency mining companies’ high-performance campuses into AI data centers has become a new trend in solving computing power bottlenecks. This long-term and significant cooperation not only provides Anthropic with long-term and stable computing power support for technological iteration, but also marks the accelerating integration of the AI industry with traditional energy infrastructure.

ENLatest AI NewsNVIDIA Enables AI to Copy Homework: KV Cache Can Be Migrated Across Models, Inference Speed Increases by 25 TimesNvidia researchers enabled KV cache migration across models, letting target models skip prefill. Conversion is 2.7–25× faster than reprocessing context. KV cache explains chatbot first-word latency: all input is preprocessed before the first word, then cache enables fast subsequent generation. This breakthrough greatly reduces response wait…..just now5.9KClaude starts secretly stamping all outputs: Anthropic launches text watermark, AI-generated content will be nowhere to hideAnthropic adds machine-readable markers to Claude outputs, with two layers: invisible text watermarks and digital signature metadata, effective globally from Aug 2. This is primarily to comply with EU AI Act transparency rules, ensuring content traceability…..just now7.9KTVB Enters the AI Computing Power Market: Launching the First Phase of Ten-Petaflop Computing Facilities in 2027TVB partners with a Gaw Capital affiliate to form a joint venture (TVB holds 51% stake) to build AI computing facilities at its Tseung Kwan O campus in Hong Kong. The project will purchase and deploy GPU hardware, offering services via a subscription model…..just now6.2KDeepSeek responds to the ‘deep thinking’ mode secretly getting a nickname: it is actually a temporary tag and will not store user informationDeepSeek clarified that the ‘secret nicknames’ from its Deep Think mode are temporary labels generated during reasoning to aid context understanding and answer optimization, not private user data logging. The issue arose from user-shared screenshots, sparking online debate and trending on social media…..just now4.3KClaire Expands Its Computing Power Map: Anthropic Collaborates with Riot to Achieve a $9.1 Billion Long-Term Cooperation AgreementThe AI computing arms race is deepening into infrastructure. Anthropic signed a 20-year, $9.1 billion deal with Riot Platforms, a bitcoin miner turned cloud service, for massive AI computing power, signaling miners’ rapid push into AI data centers…..just now7.6KTencent Opensources New Browser Automation Tool BrowserSkill: Giving AI Agents Independent Hands to WorkTencent open-sourced BrowserSkill, an AI agent browser tool designed for efficient web task handling by agents, distinct from traditional LLM plugins. Key features include full reusability, significantly boosting automation efficiency…..just now8.9KNVIDIA Collaborates with Six Financial Institutions to Raise Over $500 Billion for Building an AI Computing Financing PlatformNvidia partners with Apollo, BlackRock and four other financial institutions to create an AI infrastructure financing platform, aiming to mobilize over $500 billion in third-party capital. They’ve signed an MOU to provide funding to enterprises and cloud providers at favorable rates through specialized pools…..just now8.7KMeta Open-Sources 30B-Parameter Muse Glimmer for Local AI Agent on Consumer HardwareMeta released Muse Glimmer, a 30B-param open-source model (Apache 2.0) based on Muse Spark. Designed for local AI agents on consumer hardware, it enables multi-step tasks like tool calling, code writing/debugging, file handling, and screenshots, driving extended workflows for personal superintelligence…..just now8.9KQwen App Launches Paid Plans: Office Assistant Up to 128 Yuan/Month, AI Video Generation Charged SeparatelyAlibaba’s Qianwen App launched paid plans for pro users on Aug 11, offering office assistant subscriptions and AI video generation quotas. Three tiers: Flagship ¥128/mo (¥1499/yr) with 20x quota; Elite ¥49/mo (¥568/yr) with 5x quota; Advanced ¥19/mo (¥200/yr) with 2x quota…..just now7.2KVirgin Atlantic Uses ChatGPT Work to Redefine the Customer Journey: Weeks of Research Compressed into Hours, AI Becomes a Standard in Business OperationsVirgin Atlantic has fully integrated ChatGPT Work, helping teams accelerate market research, product planning, and decision-making. It consolidates scattered data generated throughout the customer journey, including browsing, booking, check-in, in-flight experience, and feedback, breaking down information silos and connecting data signals to help teams clarify resource allocation priorities and build a comprehensive view of the end-to-end customer journey.just now7.3KAI Promotes a Four-Day Workweek? The Reality Is That AI Efficiency Has Instead Increased the Pressure on Employees to Work 9006 (9 AM to 6 PM, Monday to Friday)Tech giants invest billions in AI, promising it will automate routine tasks, reduce work hours, and even enable a four-day work week. But ex-employees expose a starkly different reality, shattering the utopian narrative…..just now8.7KOpenAI Pays Out 7 Billion Dollars to Repurchase Employee Shares, 852 Billion Dollar Valuation Suggests IPO Is Still in the DistanceOpenAI completed a $7B employee share buyback, directly repurchasing shares without external investors, unlike past

RabbitVis – 兔展智能推出的一站式 AI 设计生产工具AI工具4天前发布AI小集03RabbitVis是什么RabbitVis 是兔展智能推出的图层级一站式AI 设计生产工具，基于自研 UniWorld-Design 视觉大模型，将 AI 生成能力与图层编辑能力深度融合。工具能通过自然语言生成海报、电商详情页、杂志内页等复杂设计内容，支持生成后的持续图层级编辑，包括去背景、文字修改、元素替换、透明素材提取等，解决传统 AI 生图只能生成、无法修改的痛点，让 AI 真正走完设计全流程。RabbitVis的主要功能智能生成：支持一句话生成海报、杂志跨页、电商详情页、信息长图、学术海报、财务分析报表等多种专业设计内容。透明素材生成：根据文本描述直接生成带透明通道（RGBA）的设计元素，可直接拖拽复用于其他设计场景。图像分层拆解：将生成的完整图像自动拆解为背景、主体、文字、装饰等多个独立可编辑图层，为后续调整提供基础。自然语言编辑：支持用自然语言指令完成去背景、图像编辑、文字编辑、橡皮擦等操作，无需掌握 Photoshop 等专业软件。中文高密度排版：针对中文场景深度优化，支持复杂艺术字体、图文交错、信息图表等高密度中文排版，效果清晰准确。如何使用RabbitVis访问官网：访问RabbitVis 官网 https://rabbitvis.rabbitpre.com/ 进入 RabbitVis 平台。描述需求：在输入框中用自然语言描述想创建的设计内容，如”生成一张618大促海报”或”设计一份杂志内页”。AI 生成初稿：系统自动生成包含完整排版、图片和文字的设计初稿，支持选择不同风格模板。图层级编辑：通过自然语言指令对设计进行持续调整，如”把 LOGO 往左移””放大标题字号””替换背景图”等。导出交付：将完成的设计导出为可交付资产，透明素材可单独下载复用。RabbitVis的核心优势图层级深度编辑：突破传统 AI 生图生成即终点的局限，支持将设计内容拆解为独立图层，实现生成后的持续修改、元素替换与文字调整。透明素材直出：根据文本描述直接生成带透明通道（RGBA）的设计元素，无需手动抠图可直接复用于多场景设计。中文场景原生优化：针对中文高密度排版深度打磨，支持复杂艺术字体、图文交错与信息图表，解决主流 AI 工具中文乱码与排版错乱痛点。自然语言驱动设计：全程通过自然语言指令完成生成、去背景、改字、调版式等操作，零设计基础用户也能产出专业级视觉物料。产模一体架构：基于自研 UniWorld-Design 视觉大模型，模型能力与产品功能深度耦合，生成、理解、编辑三位一体，越用越精准。RabbitVis的同类竞品对比对比维度RabbitVisRecraft产品定位图层级一站式 AI 设计生产工具，聚焦生成后的深度编辑与图层拆解AI 原生设计工具，侧重矢量图形生成与品牌风格一致性控制图层能力原生支持图像自动分层拆解，背景、主体、文字等元素可独立编辑调整支持基础图层管理，但主要针对矢量对象，位图分层能力有限透明素材可直接生成带透明通道（RGBA）的设计元素，语义匹配度高支持背景移除生成透明素材，但非原生透明通道生成中文排版针对中文场景深度优化，支持复杂艺术字体与高密度图文排版中文支持相对基础，复杂中文字体与竖排场景表现一般编辑方式自然语言驱动图层级编辑，零设计基础即可完成专业级修改偏向传统设计软件交互，需一定设计工具使用经验模型架构基于自研 UniWorld-Design 视觉大模型，产模一体，生成理解编辑统一基于自研 Recraft 模型，侧重矢量生成与风格迁移适用场景电商海报、杂志内页、社交媒体配图等需要反复修改的设计生产流程品牌 VI、图标库、营销插画等风格统一的设计资产创建RabbitVis的应用场景电商营销：快速生成商品详情页、促销海报、主图设计，并支持根据平台要求随时调整尺寸和文案。社交媒体运营：为公众号、小红书、抖音等平台生成配图、封面、信息长图，运营人员无需依赖设计师即可完成日常物料。品牌宣传：制作品牌手册、杂志内页、活动海报等专业视觉物料，支持持续迭代优化品牌视觉一致性。内容创作：自媒体博主、UP 主可用自然语言快速生成视频封面、专栏头图、知识图谱信息图等内容资产。企业办公：财务分析报表、PPT 配图、内部培训材料等商务视觉内容的高效生产。# 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InstructAV2AV - 智源联合北大开源的音视频联合编辑模型下一篇Muse Code - Meta 推出的终端编程 AI 智能体相关文章dinnerfy – AI餐食规划应用，根据口味偏好制定餐单，自动生成购物清单AI小集2EMO – 阿里推出的AI肖像视频生成框架AI小集23AI小方 – AI包装设计平台，自动排版配色生成创意灵感包装设计方案AI小集5GenFM – ElevenLabs推出的创建多语音AI播客功能AI小集3文佳AI – AI论文写作工具，根据论文主题和领域生成高质量写作参考AI小集3AiMakeSong – AI音乐生成平台，自动为歌词配上旋律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搜索码上飞堆友Agent美图设计室绘蛙AIupdream秒哒办公小浣熊最新收录Lawbot脱敏猫ClineJoyAI青虎AI妙呀纳逗Pro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！7分钟前LayerBack – AI 图表图片转换工具，一键转可编辑图表8分钟前2026 AI 小说写作软件最新评测：FeelFish 4.04小时前WorldClaw – 腾讯混元团队推出的 3D 世界生成框架5小时前Kitesurf – Cloudflare 推出的 AI Agent 云端轻量浏览器8小时前SmartSub – 开源的一站式音视频字幕处理桌面工具9小时前Grok Imagine Image 2.0 – SpaceXAI 推出的图像生成模型1天前Wan-Animate-2 – 万相团队开源的新一代角色动画模型1天前Alpamayo 2 Super – 英伟达开源的自动驾驶 AI 推理模型1天前Warp Agent CLI – Warp 推出的独立命令行编码工具2天前CosyVoice Studio – 阿里推出的一站式AI语音生产力平台2天前SALMONN-2 – 清华等开源的通用音频大语言模型3天前Ling-3.0-tiny – 蚂蚁百灵推出的原生混合推理模型3天前Wan3.0 – 阿里万相最新推出的视频生成大模型3天前Omega – 腾讯推出的 AI 原生数据分析平台4天前

Ling-3.0-tiny – 蚂蚁百灵推出的原生混合推理模型AI工具3天前发布AI小集02Ling-3.0-tiny是什么Ling-3.0-tiny是蚂蚁百灵推出的原生混合推理模型，总参数 7.9B、每 token 仅激活 1.3B，面向数学推理、指令遵循及资源敏感型部署优化。模型原生支持工具调用，可胜任移动设备与浏览器 UI 自动化控制任务，支持完全本地推理。目前已上线OpenRouter与 Vercel，模型权重即将开源。Ling-3.0-tiny的主要功能混合推理架构：总参数 7.9B，每 token 仅激活 1.3B，在数学推理与指令遵循任务上实现高效计算。原生工具调用：内置工具使用能力，可直接驱动浏览器 UI 控制与移动设备自动化操作。完全本地推理：无需云端依赖即可运行，集成 obsidian-cli 后可从本地文本库检索、整理并生成内容。轻量文本生成：针对本地网页翻译、日常任务处理等高频场景提供低延迟文本生成服务。Ling-3.0-tiny的技术原理稀疏混合专家架构（MoE）：模型总参数量达 7.9B，但通过门控路由机制每 token 仅激活约 1.3B 参数，在保持大模型表达能力的同时将推理计算量压缩至稠密小模型级别。原生混合推理路径：内置多条差异化推理通道，可根据任务复杂度动态选择轻量快速响应或深度链式思考路径，平衡数学推理精度与日常指令遵循的响应速度。门控路由与负载均衡：采用可学习的路由器为每个 token 分配最优专家子集，并辅以负载均衡损失函数，防止专家崩溃，确保 7.9B 参数空间被高效且均匀地利用。工具调用原生嵌入：在预训练或对齐阶段将工具使用模式直接编码进模型表征空间，使模型无需外部插件可生成符合工具接口规范的调用指令，实现浏览器 UI 与移动设备的自动化控制。端侧部署优化：针对本地推理场景进行量化压缩与 KV-Cache 内存优化，支持在消费级硬件上完全离线运行，消除云端依赖并降低高频 API 调用的网络与算力开销。如何使用Ling-3.0-tiny访问平台：访问 OpenRouter 官网（https://openrouter.ai）或 Vercel AI Gateway 页面。选择模型：搜索并进入 inclusionai/ling-3.0-tiny:free模型详情页。获取权限：注册或登录账号，在免费试用期内直接获取 API 密钥并调用服务。等待开源：关注 Ant Ling 官方渠道，待权重开源后下载模型文件进行本地部署。集成工具：配置 obsidian-cli 或浏览器自动化插件，启用本地知识检索与 UI 控制功能。Ling-3.0-tiny的核心优势极致轻量：仅激活 1.3B 参数可运行，推理成本与显存占用远低于同规模稠密模型，适合端侧与边缘部署。隐私优先：支持完全本地运行，数据不出设备，满足金融、医疗等对数据安全要求极高的场景。工具原生：无需额外微调或插件即可调用外部工具，降低自动化流程的开发门槛。Ling-3.0-tiny的同类竞品对比维度Ling-3.0-tinyQwen2.5-7B-Instruct架构原生混合推理（7.9B/1.3B 激活）稠密架构（7B 全激活）推理成本每 token 仅 1.3B 激活，极低7B 全激活，相对较高工具调用原生支持支持本地优化专为端侧与高频 API 场景深度优化通用型本地部署开源状态权重即将开源已开源Ling-3.0-tiny的应用场景移动设备智能助手：在手机端离线完成翻译、问答与日程管理，无需联网即可响应。浏览器自动化：通过原生工具调用自动执行网页表单填写、信息抓取与 UI 交互任务。本地知识管理：集成 obsidian-cli 构建完全私有的知识库，实现本地文档检索与智能摘要。企业边缘部署：部署于工控机或内网环境，在资源受限设备上提供合规的 AI 服务。高频轻量服务：作为后台常驻接口处理网页翻译、内容生成等高频请求，显著降低算力开支。# 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Wan3.0 - 阿里万相最新推出的视频生成大模型下一篇SALMONN-2 - 清华等开源的通用音频大语言模型相关文章UniReal – 港大联合 Adobe 推出的通用图像生成和编辑框架AI小集2Typeless – 智能AI语音输入工具，自动去除语气词AI小集3Project Mariner – 谷歌推出的浏览网站智能体，能帮用户操作表格、在线购物AI小集3LazyLLM – 商汤大装置开源的多智能体应用开发平台AI小集3k1 视觉思考模型 – kimi推出的 k1 系列强化学习模型AI小集4Clawra – 基于OpenClaw框架开源的AI虚拟女友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搜索码上飞堆友Agent美图设计室绘蛙AIupdream秒哒办公小浣熊最新收录Lawbot脱敏猫ClineJoyAI青虎AI妙呀纳逗Pro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LayerBack – AI 图表图片转换工具，一键转可编辑图表4分钟前2026 AI 小说写作软件最新评测：FeelFish 4.04小时前WorldClaw – 腾讯混元团队推出的 3D 世界生成框架5小时前Kitesurf – Cloudflare 推出的 AI Agent 云端轻量浏览器8小时前SmartSub – 开源的一站式音视频字幕处理桌面工具8小时前Grok Imagine Image 2.0 – SpaceXAI 推出的图像生成模型1天前Wan-Animate-2 – 万相团队开源的新一代角色动画模型1天前Alpamayo 2 Super – 英伟达开源的自动驾驶 AI 推理模型1天前Warp Agent CLI – Warp 推出的独立命令行编码工具2天前CosyVoice Studio – 阿里推出的一站式AI语音生产力平台2天前SALMONN-2 – 清华等开源的通用音频大语言模型3天前Wan3.0 – 阿里万相最新推出的视频生成大模型3天前Omega – 腾讯推出的 AI 原生数据分析平台4天前Hunyuan3D-Buffalo 1.0 – 腾讯混元推出的 3D 多模态框架4天前

Seedance 2.5 – 字节跳动推出的最新视频生成模型AI工具1周前更新AI小集03Seedance 2.5是什么Seedance 2.5 是字节跳动推出的新一代视频生成模型，单次可生成30秒高质量视频并支持多轮延长，能产出数分钟连贯叙事内容。模型支持单次输入最多30张图片、10段视频及10段音频作为参考素材，全面升级白模参考、运动参考与创意参考能力。模型新增精准时间戳编辑，支持绿幕替换、视角调整与定向修改，满足影视、广告等专业场景需求。Seedance 2.5的主要功能30秒长叙事生成：单次生成30秒高质量视频，支持多轮延长，优化镜头衔接与场景过渡，弱化”油腻感”。多模态参考输入：单次支持30张图片、10段视频、10段音频作为参考，综合理解画面构图、风格、人物与道具。白模与运动参考：支持无纹理3D白模控制空间结构、主体姿态与运镜，生成符合物理规律的光照与阴影。精准时间戳编辑：通过时间戳定向控制特定时段的故事、视角、运镜与节奏，实现生成前与生成后的精细调整。绿幕与视角编辑：替换场景时保持主体物理效果一致（衣物飘动、光影、步态），支持运镜轨迹与画面视角的独立调整。产业场景应用：覆盖教育课件生成、工业仿真、具身智能训练数据合成与自动驾驶极端场景模拟。Seedance 2.5的技术原理多模态参考融合理解：模型综合解析用户输入的30张图片、10段视频及10段音频中的画面构图、场景风格、人物特征与道具元素，通过跨模态对齐机制将参考素材中的视觉与听觉信息映射到生成目标中，实现多主体、多场景、多镜头变化的复杂创意还原。白模空间结构驱动生成：基于无纹理3D白模的空间结构、主体姿态、运动轨迹与镜头机位信息，模型构建三维空间约束，依据空间几何关系生成符合真实物理规律的光照效果，包括光源方向、色温、强度与阴影投射，使复杂镜头的构图调度与光影呈现更接近预期。时间戳精准编辑控制：模型支持时间戳级别的音视频内容定位与定向编辑，通过时序语义理解在生成阶段控制特定时段的故事走向、画面视角与运镜节奏，在生成后阶段对指定片段中的角色、动作、声音或剧情进行局部修改，同时保持修改前后的时序连贯性与真实感。绿幕物理一致性渲染：在绿幕场景替换过程中，模型基于目标场景的物理规则重新计算人物身上的物理效果，包括衣物飘动方向、头发状态、步态节奏与光影交互，确保主体与新场景在动力学与光学层面的和谐统一。如何使用Seedance 2.5选择入口：访问即梦网页端，进入视频生成模块并选择 Seedance 2.5 模型。输入创意：在输入框中描述视频内容，可上传图片、视频、音频作为多模态参考素材。延长视频：使用”延长视频”功能并指定衔接要求，保持人物主体、场景风格与声音音效一致。精准编辑：生成后通过时间戳定位特定片段，定向修改角色、动作、声音或剧情。专业控制：上传白模控制空间结构与运镜，或利用绿幕编辑替换场景背景并自动匹配物理效果。Seedance 2.5的核心优势长叙事连贯性：单次生成30秒并支持多轮延长，保持主体特征、场景环境与叙事节奏统一，减少拼接成本。多模态参考上限高：单次支持30张图、10段视频、10段音频，可同时还原多主体形象与声音，实现复杂群像叙事。精准时间戳控制：通过时间戳定向编辑特定时段的角色、动作、声音或剧情，生成前后均可精细调整，降低反复生成成本。白模物理光照：基于3D白模空间信息生成符合物理规律的光影、阴影与材质反射，提升专业影视质感。产业级编辑一致性：绿幕替换场景时保持衣物飘动方向、步态节奏、头发状态等物理效果和谐统一。Seedance 2.5的使用示例R2V prompt：延长视频，衔接 @视频 1 画面内容和主体继续生成一个 30 秒的视频，保持人物主体、场景、画面风格、声音音效一致。小男孩抱着足球沿车厢跑，地铁停稳后侧边门打开，他立刻冲出，男主紧追。两人一路穿过站台跑到街上，惊动街边路人，男主最终追上并抓住他，小男孩委屈抬头，男主慢慢消气，摸摸他的头，露出无奈笑容。R2V prompt：参考 @白模 1 的运镜、镜头节奏、景别变化、主体轨迹和镜头调度。参考 @图片 2 的角色外形、场景、材质、光照、色彩和童话氛围，将白模渲染为梦幻温暖、儿童幻想感、3D 动画短片。剧情依次为：幻想天空飞行→云海神兽伴飞→俯冲入海→鳐鱼海底穿梭→镜面时空裂缝→宇宙摘星→幻化回房间→爸爸盖被→合上绘本定格。Seedance 2.5的同类竞品对比维度Seedance 2.5可灵 3.0Runway Gen-4.5单段时长30 秒（原生）约 10-20 秒约 10-16 秒参考素材数50 个全模态有限数量有限数量局部编辑支持区域级修改部分支持支持 Inpainting分辨率原生 4K 10bit最高 1080p/4K最高 1080p价格定位国内低价策略国内中等国际高价订阅应用场景影视/广告/实体产业短视频/广告创意短片/广告Seedance 2.5的应用场景影视广告制作：模型能快速生成电影级质感短片与广告片段，支持复杂运镜、多镜头叙事与一镜到底创作。教育内容转化：将课文历史背景、科学原理转化为沉浸式视频课件，降低教学物料制作门槛。工业仿真训练：生成高质量合成视频数据，用于机器人感知操作训练与设备装配流程演示。自动驾驶测试：模拟极端天气、复杂路况等低频场景，为系统测试与训练提供多样本数据。创意内容生产：支持京剧表演、音乐会、儿童幻想动画等复杂艺术创作，实现从想法到成片全流程落地。# 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。上一篇豆包2.1 - 字节跳动推出的全新智能体模型系列下一篇方舟 CLI - 字节跳动火山引擎推出的命令行工具相关文章EmoLLM – 专注于心理健康支持的大语言模型AI小集3MicroCoder – 微软联合剑桥等推出的大模型训练优化框架AI小集3胃之书 – AI智能饮食记录应用AI小集3cogvlm2-llama3-caption – 智普AI开源的视频标注模型，生成文本描述AI小集3Qwen3.7 Preview – 阿里通义推出的下一代旗舰大模型预览版AI小集3InstantStyle – 开源的个性化文本到图像生成框架，保留风格一致性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搜索码上飞堆友Agent美图设计室绘蛙AIupdream秒哒办公小浣熊最新收录Lawbot脱敏猫ClineJoyAI青虎AI妙呀纳逗Pro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！35分钟前LayerBack – AI 图表图片转换工具，一键转可编辑图表36分钟前2026 AI 小说写作软件最新评测：FeelFish 4.05小时前WorldClaw – 腾讯混元团队推出的 3D 世界生成框架6小时前Kitesurf – Cloudflare 推出的 AI Agent 云端轻量浏览器9小时前SmartSub – 开源的一站式音视频字幕处理桌面工具9小时前Grok Imagine Image 2.0 – SpaceXAI 推出的图像生成模型1天前Wan-Animate-2 – 万相团队开源的新一代角色动画模型1天前Alpamayo 2 Super – 英伟达开源的自动驾驶 AI 推理模型1天前Warp Agent CLI – Warp 推出的独立命令行编码工具2天前CosyVoice Studio – 阿里推出的一站式AI语音生产力平台2天前SALMONN-2 – 清华等开源的通用音频大语言模型3天前Ling-3.0-tiny – 蚂蚁百灵推出的原生混合推理模型3天前Wan3.0 – 阿里万相最新推出的视频生成大模型3天前Omega – 腾讯推出的 AI 原生数据分析平台4天前

GPT-5.6 – OpenAI 推出的最新一代大语言模型系列AI工具1个月前更新AI小集04GPT-5.6是什么GPT-5.6 是 OpenAI 推出的最新一代大语言模型。模型首次采用天文学命名体系，推出三档型号：旗舰版Sol（太阳）面向高难复杂任务，是目前性能最强的版本；均衡版Terra（地球）性能对标 GPT-5.5 但价格仅一半，适合日常工作流；轻量版Luna（月亮）主打速度与低成本。核心能力聚焦于编程、生物信息学和网络安全三大方向，支持 150 万 tokens 超长上下文，在 Terminal-Bench、GeneBench 等基准测试中全面领先，新增 Ultra 模式通过多子 Agent 协作处理复杂任务。GPT-5.6的主要功能三档型号灵活选择：提供 Sol（旗舰）、Terra（均衡）、Luna（轻量）三个版本，分别对应高难度任务、日常工作和低成本高并发场景。超长上下文处理：支持 150 万 tokens 上下文窗口，可处理超长文档、代码库和复杂多轮对话。高级编程与终端操作：在命令行环境中执行复杂开发任务，支持项目结构理解、文件读取、代码修改、运行命令和报错分析。Ultra 模式多 Agent 协作：通过调度多个子 Agent 拆分处理复杂任务，统一汇总结果，提升长周期编码效率。生物信息学与基因组分析：支持长周期基因组学和定量生物分析任务，在更少输出 tokens 下获得更强结果。网络安全与漏洞研究：推进长周期安全任务，包括漏洞研究和 exploitation 相关任务，效率显著提升。分层安全防护体系：加强对高风险活动、敏感网络请求和重复滥用行为的防护，抵御真实世界攻击。提示词缓存与成本优化：优化重复提示词调用机制，降低高频场景下的 API 使用成本。如何使用GPT-5.6ChatGPT：Plus/Pro/Business/Enterprise用户可访问Sol（medium及以上effort），Pro和Enterprise可选Sol Pro。ChatGPT Work/Codex：免费和Go用户用Terra，Plus及以上可选Sol/Terra/Luna，可设置effort等级；max模式所有可用用户可开启；ultra在ChatGPT Work中Pro/Enterprise可用，Codex中Plus及以上可用。API：通过OpenAI API访问三档模型，Responses API支持Programmatic Tool Calling（ZDR兼容）和Multi-agent beta。GPT-5.6的官网地址官网地址：https://openai.com/index/gpt-5-6/GPT-5.6的核心优势三档分级灵活匹配：Sol（旗舰）、Terra（均衡）、Luna（轻量）覆盖高难度任务、日常工作和低成本高并发场景，按需选择不浪费算力。编程能力行业领先：Terminal-Bench 2.1 得分 Sol 88.8%、Sol Ultra 91.9%，超越 Claude Mythos 5（88.0%）等竞品。超长上下文窗口：支持 150 万 tokens，比 GPT-5.5 提升约 50%，可处理完整代码库、长篇文献和复杂多轮对话。Ultra 模式多 Agent 协作：自动调度多个子 Agent 拆分并行处理复杂任务，再统一汇总结果，大幅提升长周期编码效率。科研成本效率双优：GeneBench v1 上 Sol 比 GPT-5.5 得分更高，但输出 tokens 更少，生物信息学分析更省成本。网络安全高效精准：ExploitBench 接近 Mythos Preview 水平，但仅使用约 1/3 的输出 tokens，漏洞研究效率显著提升。分层安全防护体系：迄今最强安全堆栈，可抵御高风险活动、敏感网络请求和真实世界攻击。价格梯度合理：Terra 性能对标 GPT-5.5 但价格仅一半；Luna 成本最低；提示词缓存机制进一步降低重复调用开销。GPT-5.6的同类竞品对比对比维度GPT-5.6 Sol (OpenAI)Claude Fable 5(Anthropic)Gemini 3.1 ProPreview (Google)发布时间2026年6月27日2026年6月9日2026年2月19日模型定位旗舰版（三档分级：Sol/Terra/Luna）Mythos-class（高于Opus级别）旗舰多模态模型上下文窗口150万 tokens100万 tokens100万 tokensTerminal-Bench 2.188.8%84.3%70.7%SWE-bench Verified未公开具体分数95.0%未公开具体分数输入价格（/百万tokens）$5$10$2输出价格（/百万tokens）$30$50$12多模态能力文本、图像文本、图像、PDF文本、图像、音频、视频、PDF安全策略分层安全堆栈，需政府审批访问实时安全分类器，敏感查询自动降级至Opus 4.8标准安全策略GPT-5.6的应用场景复杂软件开发与终端操作：深入命令行环境执行复杂编码任务，包括项目结构理解、文件读取、代码修改、运行命令、报错分析与迭代修复，适合大型代码库重构和全栈开发。生物信息学与基因组研究：处理长周期基因组学分析和定量生物分析任务，在多轮数据解释、方法选择和假设比较中保持上下文一致，适合实验室和生物医药研发团队。网络安全与漏洞研究：执行长周期安全任务，包括漏洞挖掘、exploitation 研究和渗透测试，以更低 token 消耗实现高效分析。企业日常办公自动化：通过 Terra 版本处理文档分析、报告撰写、数据整理等高频日常工作流，性能对标 GPT-5.5 但成本减半。高并发低成本推理服务：通过 Luna 版本承载客服机器人、内容摘要、简单问答等大规模高容量调用场景，以最低价格提供强大基础能力。超长文档深度分析：用 150 万 tokens 上下文窗口处理完整技术手册、法律合同、学术论文或大型代码库的全文理解与跨章节关联分析。# 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Papers with Code - AI科研平台，自动关联论文、代码等下一篇HTML Anything - nexu-io 团队开源的 HTML 编辑器相关文章HoloPart – 港大联合 VAST 开源生成完整可编辑部件的 3D 模型AI小集3SearchClaw – 中国人民大学推出的 AI 深度研究智能体AI小集3CombatVLA – 淘天集团推出的3D动作游戏专用VLA模型AI小集2SoulX-Singer – Soul App联合高校开源的歌声合成模型AI小集3Fineshare VoiceTrans – 实时AI语音转换器，实时声音调节保留口音、语调、情感AI小集2点线搜索 – 零一万物推出的AI搜索应用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搜索码上飞堆友Agent美图设计室绘蛙AIupdream秒哒办公小浣熊最新收录Lawbot脱敏猫ClineJoyAI青虎AI妙呀纳逗Pro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！5分钟前LayerBack – AI 图表图片转换工具，一键转可编辑图表6分钟前2026 AI 小说写作软件最新评测：FeelFish 4.04小时前WorldClaw – 腾讯混元团队推出的 3D 世界生成框架5小时前Kitesurf – Cloudflare 推出的 AI Agent 云端轻量浏览器8小时前SmartSub – 开源的一站式音视频字幕处理桌面工具9小时前Grok Imagine Image 2.0 – SpaceXAI 推出的图像生成模型1天前Wan-Animate-2 – 万相团队开源的新一代角色动画模型1天前Alpamayo 2 Super – 英伟达开源的自动驾驶 AI 推理模型1天前Warp Agent CLI – Warp 推出的独立命令行编码工具2天前CosyVoice Studio – 阿里推出的一站式AI语音生产力平台2天前SALMONN-2 – 清华等开源的通用音频大语言模型3天前Ling-3.0-tiny – 蚂蚁百灵推出的原生混合推理模型3天前Wan3.0 – 阿里万相最新推出的视频生成大模型3天前Omega – 腾讯推出的 AI 原生数据分析平台4天前