AI NEWSLatest AI NewsArticleWorkBuddy Major Update: ‘Human-Machine Dual Writing’ Launches, the Office of the AI Era Has Arrived!Published in Latest AI NewsTime :Jul 30, 2026Read :9minuteOn July 30, Tencent WorkBuddy released a major version update (V5.3.5): in collaboration with Tencent Docs, it officially launched the “human-AI dual-writing” collaborative editing capability for the Agent era. The new version allows users to directly collaborate with AI within the same file when opening Word, Excel, PPT, Markdown, and other document formats, addressing pain points such as multi-role collaboration in “human-AI hybrid writing” teams.From the traditional “human-file interaction” of conventional Office, to “human-human collaboration” in online documents, to “human-AI task completion” in the AI era, the launch of this feature in WorkBuddy has made AI the first time an “add-on” in office collaboration become a collaborator within the same document. It also achieved, for the first time in the industry, “human-human, human-AI, AI-AI” multi-end synchronized collaboration, driving Office into an AI-native collaboration era.AI Enters Documents: From “Help You Write” to “Write with You"Previously, users needed to switch between chat windows and Office software when using AI for work. To get AI to generate content, they had to copy and paste it into the document; to ask AI to modify existing materials, they had to re-describe the background and revise the entire text; when working in teams, it was difficult to collaborate in real-time around the same document with the same AI partner. Although AI could generate content, it was hard to truly integrate into the user’s ongoing workflow.With this upgrade, WorkBuddy now directly integrates AI capabilities into the document editing environment. After opening a Tencent Docs or local Office file in WorkBuddy, users can select text, table areas, or PPT pages in the original file and use natural language to tell AI what task needs to be completed. For example, selecting a market analysis report to optimize its expression; circling a set of Excel data to have AI analyze it and generate charts; specifying a PPT page to have AI adjust its structure and content; marking blank areas in a report to have AI add relevant sections.AI-generated content appears directly in the original file, where users can continue to edit or add comments. AI will perceive and complete the modifications in real time. After AI completes the edits, users can take over and continue writing at any time. This means that AI is no longer just a “side panel assistant,” but becomes a true collaborator in the creation process.Three Core Upgrades Make AI Transition from “Auxiliary Add-on” to “Third Colleague"The core of this upgrade is to give AI more document operation capabilities and context awareness similar to real collaborators, which are mainly reflected in three aspects:First, it covers more files and scenarios, supporting full human-AI co-writing.Relying on Tencent Docs’ editing capabilities, the latest version of WorkBuddy supports mainstream office formats such as Markdown, Word, Excel, and PPT, and covers both local Office files and online Tencent Docs. Whether it is a thesis, proposal, lesson plan, technical documentation, consulting report, or operational weekly reports, pitch materials, and data analysis, users can complete creation, editing, and collaboration in the same workflow.Second, selecting and modifying is precise and efficient.Users can select paragraphs, data areas, or PPT canvases and let AI make changes only within the specified range, keeping the rest of the layout and content unchanged. For example, in Excel, you can circle thousands of rows of data and let AI perform a pivot analysis, generate trend charts, or batch correct errors and apply complex formulas. Switching to PPT, you can let AI automatically layout the analysis conclusions and charts in a blank area. For structured documents like proposals and lesson plans, you can select the areas to be filled and let AI complete them according to the template.Third, multiple people plus AI, real-time collaboration without interruption.AI-generated content can be uploaded to Tencent Docs with one click, allowing team members to jointly edit the same document across different terminals, as well as collaborate with the same AI partner. All changes are synchronized in real time, and the document context remains intact. From personal writing to team collaboration, to multiple people working on the same task with AI together, human collaboration, human-AI collaboration, and AI-AI collaboration are seamlessly integrated within the same document.From the first generation solving individual local office work, to the second generation solving multi-person collaboration through online documents, as AI agents are deeply integrated into enterprise productivity, office software is moving towards the third stage. By enabling “human-human, human-AI, AI-A

AI NEWSLatest AI NewsArticleConcerns about Microsoft’s Cloud: Growth Deceleration and Leverage Risks Behind Billions in RevenuePublished in Latest AI NewsTime :Jul 31, 2026Read :3minuteIn today’s cloud services market, the arms race among major players is accelerating rapidly. Microsoftachieved its first annual revenue of over $100 billion in the just-completed fiscal year, with a year-over-year growth of 43%. However, this performance does not guarantee peace of mind, as its business growth faces challenges and is even surpassed bywith an astonishing growth rate of 82%.The reason behind this lies in the different paths both giants have chosen in terms of technical approaches and business returns. Google, with its complete ecosystem of self-developed models and chips, has significantly increased its cloud operating profit margin from 20.7% to 35.6%, doubling its profits. In contrast, Microsoft, whose Maia and Cobalt chips are still in early stages, currently relies heavily on commercial chips from NVIDIA. The reselling stack model not only thins its profit structure but also puts more pressure on its computing cost.At the same time, Microsoft’s large order backlog has drawn deep scrutiny from the market. A significant portion of its $67.8 billion in contract backlogs depends on continuous investment from a single client. This highly concentrated client structure and the cycle of capital market lending have raised concerns about potential risks in infrastructure.Against the backdrop of intensifying industry competition, major cloud providers continue to increase their investments. How to find a balance between aggressive computing power arms race and stable financial health will become key to reshaping the future cloud landscape.

AI NEWSLatest AI NewsArticleWorld Bank tells emerging economies: Don’t compete with rich countries on data centers and large models, local small tools are the way to goPublished in Latest AI NewsTime :Aug 5, 2026Read :4minuteThe World Bank has outlined a clear path for developing economies in its latest report: adopt artificial intelligence into government governance and business operations as soon as possible, and adapt existing tools to local needs—but there’s no need to compete with wealthy countries in building large data centers or large language models. Instead of investing heavily in computing power, promoting small, low-cost AI tools that have been locally adapted is more likely to truly improve the quality of healthcare and education.This call to action is backed by a sobering reality. Developing economies are currently experiencing the weakest average growth performance in 30 years, and the World Bank believes AI has the potential to bring them back on track for growth. “AI is more likely to be a helper for workers rather than taking their jobs,” the report states. It envisions a compressed timeline: through AI, hundreds of millions of people who have long lacked access to services can obtain expensive medical, legal, educational, and agricultural services. Progress that might have taken 100 years in the past could now be achieved within 10 years.The impact of job displacement will not be evenly distributed. The risk of automation taking jobs in high-income countries is more than three times that of middle- and low-income countries; conversely, AI can also help alleviate shortages of skilled professionals, assisting teachers with lesson planning, helping nurses interpret medical images, and giving farmers planting advice, thereby boosting human efficiency.However, the real obstacle lies in the foundation. Some developing countries still lack internet access, electricity, and AI skills—by 2024, three out of every 10 rural schools in sub-Saharan Africa lack stable electricity, and 89% of 10-year-old children cannot read simple texts. The World Bank warns: if infrastructure, talent, institutions, and financing conditions do not improve, AI will not narrow the gap but may even widen it. The report also highlights concerns: jobs in call centers and basic software outsourcing may decline, and social inequality, cybercrime, and dependence on foreign technology may increase as a result.

AI NEWSLatest AI NewsArticleTencent Hunyuan Hy ASR3.0 Preview Release: From Hearing to Understanding, Accents and Word Error Rate Reduced to 3%Published in Latest AI NewsTime :Aug 5, 2026Read :4minuteTencent Hunyuan has officially launched the new generation speech recognition model Hy ASR3.0 Preview, marking a new stage in speech recognition technology from “word-by-word transcription” to “contextual understanding.” The model is built on the latest generation large language model Hy3, integrating high-precision speech recognition and deep semantic understanding capabilities, with the core goal of achieving a qualitative leap—making AI not only clearly hear each word but truly understand what you are saying.Supports ten dialect regions, with outstanding performance in long audio scenariosIn terms of coverage, Hy ASR3.0 Preview does not require standard Mandarin. It supports 10 major dialect regions, including Cantonese and Wu, and more than 20 secondary sub-regions. Open-source evaluation data shows that the model’s word error rate (WER) for multiple languages is controlled at around 3%, with Mandarin at 3.34%, English at 2.62%, and Cantonese at 3.12%. The overall accuracy is already close to the industry ceiling.Combined with Hy3’s language understanding capabilities, the model can accurately capture user intent by considering the context, automatically eliminate ambiguities, and intelligently correct homophones. In scenarios such as meeting minutes and long audio interviews, this “understand first, then transcribe” approach shows significant advantages—traditional word-by-word transcription often struggles with homophonic ambiguities, while Hy ASR3.0 can automatically select the correct words based on semantic logic.MoE architecture as the foundation, trained on millions of hours of dataOn the technical level, the model uses an MoE architecture that balances efficiency and performance, with the base upgraded to Hy3. The Hunyuan team developed a self-researched unsupervised speech Encoder, which was trained on tens of millions of hours of unsupervised speech data to extract high-quality acoustic representations from complex audio—Encoder is responsible for “hearing well,” while Hy3 is responsible for “understanding correctly.“In terms of data strategy, the team conducted joint training of the speech Encoder and large language model, introducing tens of millions of hours of multi-source speech data, and through multi-stage capability injection, enabling the model to have context awareness, adaptability to complex scenarios, and dialect recognition abilities. The post-training phase focuses on building a high-quality SFT solution covering general recognition, context understanding, and robustness in complex scenarios, covering context, specialized terms, different acoustic environments, and diverse voice inputs. It also introduces a multi-stage reinforcement learning approach to continuously optimize complex long-tail scenarios.

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.

ENLatest AI NewsNo Cloud API Required! Liquid AI Launches Local-Running LFM2.5 Edge Intelligence ModelOn August 5, Liquid AI launched LFM2.5-2.6B, a 2.6B-parameter on-device model for agentic workflows, running fully locally on phones with low cost, low latency, and privacy. Pre-trained on 34T tokens with 128K vocabulary, it uses supervised fine-tuning, teacher specialization, and multi-domain strategies…..7 hours ago60.8KTencent Hunyuan Hy ASR3.0 Preview Release: From Hearing to Understanding, Accents and Word Error Rate Reduced to 3%Tencent Hunyuan launches Hy ASR3.0 Preview, built on the large language model Hy3. It integrates high-accuracy recognition with deep semantic understanding, enabling a shift from word-by-word transcription to contextual comprehension. Covering ten major dialect regions, it excels in long audio scenarios without relying on standard Mandarin, allowing AI to truly understand speech…..7 hours ago12.7KJi Meng AI Integrates with Seedance 2.5 and Launches Maya/Blender Plugins, ByteDance Completes the Full Workflow of AI Video Professional CreationByteDance’s Jimeng AI integrates Seedance 2.5 model, launches Maya & Blender plugins, accelerating shift to pro film/3D creation. Single video generation length doubled to 30s…..8 hours ago12.2KWorld Bank tells emerging economies: Don’t compete with rich countries on data centers and large models, local small tools are the way to goWorld Bank report advises developing economies to swiftly embed AI in governance and business, adapting tools to local needs instead of competing with wealthy nations on large models and data centers. Promoting localized, low-cost AI can truly improve healthcare and education amid the weakest growth in 30 years…..8 hours ago10.7KWuwen Xinqiong Partners with MiniMax: From Inference Efficiency to Token Services, Domestic Large Models Join Forces to Build an EcosystemWuwen Xinqiong and MiniMax signed a strategic partnership focusing on four areas: jointly optimizing model inference efficiency, building large model token services and AI infrastructure, co-innovating application paradigms, and expanding industry intelligent scenarios. They complement each other: Wuwen excels in computing power scheduling and inference infrastructure, while MiniMax has self-developed large models and massive application traffic…..8 hours ago12.7KMistral Launches Shieldstral: A 3B Small Model Runs on a Single 16GB GPU for Multimodal Moderation, Claims to Achieve Open-Source SOTAMistral AI released the open-source content moderation model Shieldstral, with 3B parameters, under the Apache 2.0 license, supporting 12 languages and running on a single 16GB GPU. It claims performance comparable to 7 times larger models and achieves SOTA in multimodal moderation, but points out that most protection models have a fixed harm classification system.8 hours ago12.0KRecord-breaking! A total of 8 Pulitzer Prize winners and finalists have reported using artificial intelligence this yearThis year, a total of 8 Pulitzer Prize winners and finalists have voluntarily disclosed the use of AI, setting a record since the introduction of AI reporting rules last year, with 5 of them ultimately winning. The report shows that this year’s participants have more frequently used generative AI and large models for basic newsroom tasks such as searching through massive documents, organizing data, and cross-verifying information.9 hours ago14.0KVisa Spends 2.4 Billion Dollars to Acquire AI Fraud Prevention Company BioCatch: Analyzing 19 Billion Interactions Per Month to Curb Financial Fraud at the SourcePayment giant Visa announced a $2.4 billion cash acquisition of Israeli AI fraud prevention company BioCatch, integrating its behavioral analysis and device intelligence technologies to enhance risk detection capabilities for account theft, fraud, fake accounts, and false account openings. The transaction is subject to regulatory approval and is expected to close by the end of March 2027. BioCatch was founded in 2011 and specializes in AI fraud prevention.9 hours ago11.7K52,000 Yuan upon Signing! Robot Unicorn Yushu Technology Advances to the Sci-Tech Innovation Board, Estimated Issue Price Exceeds 100 YuanYushu Technology started the preliminary inquiry for its Sci-Tech Innovation Board IPO on August 5th, determined the issue price and valid quotation on the 6th, started the subscription on the 10th, and the payment deadline is on the 12th. The company plans to issue 40.4464 million new shares.10 hours ago12.4KByteDance Seed Launches SeedRealtime: Full-duplex Audio-Visual Large Model for Car Integration, Doubao Allows Watching, Listening, and Speaking Without LagByteDance’s Seed team launched SeedRealtime, a unified model that natively fuses audio, video, and text for full-duplex real-time interaction—enabling seeing, listening, and speaking simultaneously. It is now fully deployed in the Doubao App, marking the first large-scale implementation. Unlik

SenseNova U1.5-Lite-Preview – 商汤开源的轻量多模态模型AI工具2天前发布AI小集02SenseNova U1.5-Lite-Preview是什么SenseNova U1.5-Lite-Preview 是商汤科技开源的轻量级原生统一多模态模型预览版，基于 NEO-Unify 架构迭代，仅 8B-MoT 参数即贯通视觉理解、推理、生成与编辑。模型原生支持 4K 图像生成，具备精细局部纹理、真实世界质感、中英文文字生成与复杂版式能力，优化对超长自然语言和结构化视觉指令的精准遵循，配套 Prompt Enhance Skill 降低复杂创作门槛。SenseNova U1.5-Lite-Preview的主要功能4K 原生生成：支持超高分辨率图像端到端生成，细节经得起放大。精细视觉质感：呈现更真实的材质纹理、光影层次与局部细节。中英文文字生成：准确渲染复杂版式中的中英文文本与排版。图像编辑与迭代：支持基于自然语言指令的可持续迭代图像编辑。Prompt Enhance Skill：自动将简短需求扩展为完整创作方案，降低指令编写门槛。SenseNova U1.5-Lite-Preview的技术原理NEO-Unify 统一架构：在单一模型中联合建模语言、视觉语义与像素生成，实现理解、推理、生成与编辑的原生统一。8B-MoT 轻量设计：采用仅 80 亿参数的混合 token 架构，在保持轻量化的同时扩展多模态能力边界，验证统一架构的持续可扩展性。原生像素-语言映射：从像素和语言出发端到端训练，模型学到的是从语言描述到视觉结构的原生映射能力。长上下文视觉控制：模型优化对超长自然语言和结构化创作指令的理解，支持多约束、层次化视觉指令的精准执行。微信关注回复“开源”，加入AI开源项目交流群如何使用SenseNova U1.5-Lite-Preview获取模型与文档：访问 GitHub 仓库查看技术文档与推理代码，或在 Hugging Face 下载 8B-MoT 模型权重与配置文件。环境部署：依据文档配置运行环境并加载模型权重，完成本地或云端的模型初始化。基础图像生成：直接输入自然语言描述（支持中英文），模型即可原生生成 4K 图像。复杂创作指令：提供包含主体、布局、风格、文字及约束条件的完整创作要求，模型按明确视觉结构执行生成。Prompt Enhance 辅助：调用实验性 Prompt Enhance Skill 将简短想法自动扩展为完整创作方案，再提交模型执行。图像编辑操作：上传已有图像并附加自然语言编辑指令，模型在保持整体一致性的前提下完成可持续迭代编辑。SenseNova U1.5-Lite-Preview的核心优势轻量统一：仅 8B-MoT 参数即实现理解、生成、编辑一体化，降低部署成本。4K 超高清：原生支持 4K 分辨率生成，超大画幅下仍保持细节与一致性。强指令遵循：在无专门 Prompt Enhance 格式化训练下，仍能稳定执行长篇复杂视觉指令。真实质感：显著提升局部纹理、材质光影与真实世界质感表现。文字稳定性：中英文文字生成与复杂版式组织更加准确稳定。SenseNova U1.5-Lite-Preview的项目地址GitHub仓库：https://github.com/OpenSenseNova/SenseNova-U1/blob/main/docs/u1.5_preview.mdHuggingFace模型库：https://huggingface.co/sensenova/SenseNova-U1.5-8B-MoT-PreviewSenseNova U1.5-Lite-Preview的同类竞品对比维度SenseNova U1.5-Lite-PreviewEmu3（智源研究院）架构路线NEO-Unify 原生统一多模态，理解/推理/生成/编辑一体化原生统一多模态，自回归框架下统一图像理解与生成开源规模8B-MoT 轻量开源，面向社区预览提供多尺寸开源版本（如 8B 等），训练代码与数据流程完整分辨率支持原生支持 4K 图像生成，细节与一致性突出支持高分辨率生成，主要聚焦常规尺寸与视频帧生成核心能力超长指令遵循、4K 超高清、中英文文字渲染、可持续迭代编辑图像理解、文本到图像生成、视频预测，强调多模态统一文字与版式针对中英文文字生成与复杂版式进行专项优化，稳定性高文字生成能力存在，但版式控制与长文本准确性相对有限定位轻量级统一多模态基座预览版，侧重视觉创作与编辑开源统一多模态基础模型，侧重研究验证与多模态统一范式SenseNova U1.5-Lite-Preview的应用场景商业品牌视觉设计：快速生成包含精确中英文文字、复杂版式与指定视觉风格的海报、信息图及品牌宣传物料，满足高控制上限的商业创作需求。超宽幅叙事长卷：用 4K 超高清分辨率创作历史年表、文化长卷等超宽幅连续叙事内容，大量文字与图标细节在放大后依然清晰稳定。建筑与产品概念可视化：产出高真实感的建筑外观渲染、室内空间概念图及产品展示图，呈现真实的材质纹理、光影层次与空间氛围。摄影级人像与特写：模型能生成具有精细皮肤纹理、自然光影与真实质感的人像摄影、面部特写及人物主题作品，支持复杂构图与风格指定。复古与拼贴艺术创作：模型支持多元素融合的复古手帐、拼贴海报、博物学图鉴等风格化视觉创作，准确呈现纸张质感、邮票、手写文字等细节。# 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Polar - AI 浏览器，自动跨站收集整合信息下一篇E-Bench - 腾讯混元等推出的智能体评测基准相关文章MineWorld – 微软研究院开源的实时交互式世界模型AI小集2Seed1.6 – 字节跳动推出的通用模型系列AI小集3MaineCoon – AI 实时音视频世界模型，专为社交互动场景AI小集2Step-Video-T2V – 阶跃星辰开源的文本到视频模型AI小集3KAZE – 在线AI图片处理平台，精准识别清除水印区域AI小集3sCM – OpenAI推出连续时间一致性模型，两步采样生成高质量图像AI小集3暂无评论再想想发表评论暂无评论…热门工具LoomyLibTV绘蛙AIAiPPT秘塔AI搜索蛙蛙写作堆友Agent美图设计室Sekoupdream秒哒办公小浣熊最新收录青虎AI妙呀纳逗Pro图改改AstraFlow星图不繁简历最新文章秒悟Meoo – 1分钟生成前后端完整的网站，真0门槛开发！2小时前Get3W – AI 模型聚合平台，统一接口自由切换模型2小时前Shieldstral – Mistral AI 开源的多模态内容安全分类模型2小时前MAGI-2-preview – Sand.ai 开源的多模态视频生成模型2小时前SeedRealtime – 字节跳动推出的原生音视频全双工大模型2小时前JoyAI-Video-Edit – 京东开源的实时流式视频编辑模型2小时前Hy ASR 3.0 preview – 腾讯混元推出的新一代语音识别模型9小时前Qwen-CUA – 阿里 Qwen 等推出的原生 Computer Use Agent10小时前MemHarness – 上海 AI Lab 等推出的智能体记忆重构框架11小时前ForgeStencil – 面壁智能推出的 Stencil 全自动研究部署系统1天前Orchard – 微软研究院开源的 Agentic AI 建模框架1天前E-Bench – 腾讯混元等推出的智能体评测基准1天前Polar – AI 浏览器，自动跨站收集整合信息2天前GenOffice – Genspark 推出的本地 AI Office 客户端2天前Qwen 3.8-Max – 阿里云 Qwen 团队推出的旗舰大模型2天前

JoyAI-Video-Edit – 京东开源的实时流式视频编辑模型AI工具2小时前发布AI小集02JoyAI-Video-Edit是什么JoyAI-Video-Edit是京东开源自研的实时流式视频编辑模型，基于16B参数自回归扩散架构，在720P分辨率下实现30FPS推理速度，支持任意时长稳定流式编辑。用户可通过自然语言指令在视频播放过程中实时修改人物、场景、风格及物体，无需等待完整视频生成。模型在OpenVE-Bench评测中超越所有流式编辑方法，各项指标全面领先，同时可为具身智能提供大规模数据合成路径。JoyAI-Video-Edit的主要功能实时流式编辑：视频帧到达时即时编辑，无需预先知道完整序列。开放式指令控制：支持全局风格迁移、局部对象增删改、背景替换、动作调整及参考图引导编辑。任意时长处理：突破秒级或分钟级限制，可持续随视频播放进行稳定编辑。720P高吞吐部署：端到端pipeline在720×1280分辨率下达30.19 FPS。多维度编辑能力：覆盖人物换装、场景变换、风格转换、物体替换与字幕编辑等任务。JoyAI-Video-Edit的技术原理MLLM条件编码器：用多模态大语言模型将自然语言指令编码为编辑条件，实现开放式语义控制。因果视频VAE：采用因果时序结构的变分自编码器对视频帧进行压缩与重建，保证流式场景下的时序一致性。16B参数MMDiT骨干：用16B多模态扩散Transformer为核心，融合文本与视觉特征进行自回归扩散生成。自回归分布匹配蒸馏：通过对齐自回归分布的蒸馏策略加速推理，显著降低训练与推理的分布差异。长时域优化与有界KV推理：引入长时域稳定性优化及有界KV缓存机制，抑制累积时序漂移并维持高吞吐量。微信关注回复“开源”，加入AI开源项目交流群如何使用JoyAI-Video-Edit环境准备：创建Conda环境python=3.10并安装requirements.txt依赖。下载权重：从Hugging Face获取模型文件，按指定目录结构放置检查点。启动服务：进入deploy目录执行bash run_server.sh启动本地服务器。访问界面：浏览器打开http://localhost:8080进入交互式编辑界面。实时编辑：上传视频或接入摄像头流，输入自然语言指令即可边播边改。JoyAI-Video-Edit的核心优势实时流式编辑：视频帧到达即处理，无需等待完整序列，实现真正的”边播边改”交互体验。速度与质量双优：720P分辨率下端到端推理达30.19 FPS，在OpenVE-Bench评测中超越所有流式与离线编辑方法。任意时长稳定输出：突破传统模型仅支持秒级或分钟级片段的限制，可持续随视频播放进行无漂移稳定编辑。开放式语义控制：基于MLLM条件编码器，支持自然语言指令驱动的全局风格、局部对象、背景替换及参考图引导等多维度编辑。高效推理架构：16B参数自回归扩散Transformer配合分布匹配蒸馏与有界KV缓存，显著降低训练推理差异并抑制时序累积误差。JoyAI-Video-Edit的项目地址GitHub仓库：https://github.com/jd-opensource/JoyAI-Video-EditHuggingFace模型库：https://huggingface.co/jdopensource/JoyAI-Video-EditarXiv技术论文：https://arxiv.org/pdf/2608.03974JoyAI-Video-Edit的同类竞品对比维度JoyAI-Video-EditSANA-Streaming模型规模16B参数2B参数推理速度37.21 FPS (DiT)约54 FPS (DiT)编辑质量3.60 (OpenVE-Bench)约2.6编辑类型开放式多维度编辑流式编辑分辨率支持720P稳定输出较低分辨率为主时序稳定性任意时长，有界KV优化短片段为主JoyAI-Video-Edit的应用场景短视频创作：创作者可在视频播放过程中实时替换人物服装、调整场景风格，提升内容生产效率。直播实时特效：主播边播边改背景或添加虚拟元素，实现互动式直播视觉增强。家装与室内设计：用户在浏览房间视频时实时更换家具、墙面材质与灯光效果，辅助装修决策。影视后期预演：导演在剪辑阶段快速尝试不同视觉风格与场景替换，降低试错成本。具身智能数据合成：将人手操作视频转化为机械臂操作素材，低成本扩展机器人训练数据规模。# 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Hy ASR 3.0 preview - 腾讯混元推出的新一代语音识别模型下一篇SeedRealtime - 字节跳动推出的原生音视频全双工大模型相关文章Desearch – AI深度研究工具，自动梳理研究思路生成研究框架AI小集3Gemini CLI – 谷歌开源的AI编程工具，免费用Gemini 2.5 Pro模型AI小集3yaara.ai – AI写作辅助工具，支持25+种语言多风格写作AI小集2FakeShield – 北大推出检测图像伪造的多模态大语言模型框架AI小集3MoBA – Moonshot AI 提出的新型注意力机制AI小集3Animate Anyone 2 – 阿里通义推出的高保真角色图像动画生成技术AI小集3暂无评论再想想发表评论暂无评论…热门工具LoomyLibTV绘蛙AIAiPPT秘塔AI搜索蛙蛙写作堆友Agent美图设计室Sekoupdream秒哒办公小浣熊最新收录青虎AI妙呀纳逗Pro图改改AstraFlow星图不繁简历最新文章秒悟Meoo – 1分钟生成前后端完整的网站，真0门槛开发！1小时前Get3W – AI 模型聚合平台，统一接口自由切换模型1小时前Shieldstral – Mistral AI 开源的多模态内容安全分类模型2小时前MAGI-2-preview – Sand.ai 开源的多模态视频生成模型2小时前SeedRealtime – 字节跳动推出的原生音视频全双工大模型2小时前Hy ASR 3.0 preview – 腾讯混元推出的新一代语音识别模型9小时前Qwen-CUA – 阿里 Qwen 等推出的原生 Computer Use Agent10小时前MemHarness – 上海 AI Lab 等推出的智能体记忆重构框架10小时前ForgeStencil – 面壁智能推出的 Stencil 全自动研究部署系统1天前Orchard – 微软研究院开源的 Agentic AI 建模框架1天前E-Bench – 腾讯混元等推出的智能体评测基准1天前SenseNova U1.5-Lite-Preview – 商汤开源的轻量多模态模型2天前Polar – AI 浏览器，自动跨站收集整合信息2天前GenOffice – Genspark 推出的本地 AI Office 客户端2天前Qwen 3.8-Max – 阿里云 Qwen 团队推出的旗舰大模型2天前

Shieldstral – Mistral AI 开源的多模态内容安全分类模型AI工具2小时前发布AI小集02Shieldstral是什么Shieldstral 是 Mistral AI 推出的 3B 参数开源多模态内容安全分类模型，基于Ministral-3B构建。模型将传统固定类别的内容审核重新定义为二元问答任务，支持通过自然语言查询实时定义审核策略，无需重新训练即可适配不同场景。模型在文本安全基准上平均 F1 达 84.9%，多模态安全 F1 达 83.8% 均为 SOTA，且仅需单张 16GB GPU 可部署，支持 12 种语言。Shieldstral的主要功能策略自适应审核：将安全政策以自然语言问题形式输入，实现不同场景下的实时定制化审核。多模态统一检测：单一接口同时处理纯文本、图像及图文混合内容的安全评估。细粒度违规识别：支持从粗粒度二元判断到 73 个叶类别的精细安全分类。校准安全评分：通过 softmax 归一化输出 0–1 区间的连续安全分数，以 0.5 阈值判定违规。轻量端侧部署：3B 参数规模可在单张 16GB NVIDIA GPU 上高效推理，降低硬件门槛。Shieldstral的技术原理二元问答架构：将审核任务转化为对 “yes” 与 “no” 两个输出词元的逻辑值预测，经 softmax 归一化得到连续安全分数。三字段结构化输入：采用（评估场景与严格度）、（是/否安全问题）、（待审核内容）的标准化提示格式。大规模对比训练：基于 5410 万样本（4520 万开源文本、440 万合成对比文本、450 万多模态），通过同内容异查询的对比配对训练模型区分相似类别的能力。合成数据增强：利用 LLM 将安全文本改写为违规变体，并自动生成目标类别与兄弟类别的对比查询对，强化细粒度判别。LoRA 微调 + SLERP 合并：对 Ministral-3B 进行 LoRA 高效微调，并通过球面线性插值合并公共数据集与合成数据训练的两个互补检查点。微信关注回复“开源”，加入AI开源项目交流群如何使用Shieldstral环境准备：在单张 16GB NVIDIA GPU 上部署 Shieldstral 模型及推理框架。定义审核策略：编写字段说明评估场景、领域背景与严格程度标准。编写安全查询：构造字段中的二元是/否问题，例如”内容是否宣扬身体暴力？”。输入待审内容：将文本、图像或图文组合填入字段提交模型。获取安全判定：读取模型输出的 softmax 归一化连续分数，按业务需求设定阈值进行二元违规判定。Shieldstral的核心优势策略灵活：审核标准通过自然语言实时定义，无需针对新场景重新训练模型。性能领先：3B 参数在文本和多模态安全基准上均达到 SOTA，媲美 7 倍规模模型。硬件友好：单张 16GB GPU 即可运行，显著降低企业私有化部署门槛。数据统一：通过指令-查询-文档格式整合 54.1M 异构数据，覆盖 12 种语言与多元场景。决策可解释：输出校准后的连续分数而非黑盒分类标签，便于运营人员按需调整阈值。Shieldstral的项目地址项目官网：https://mistral.ai/news/shieldstral/HuggingFace模型库：https://huggingface.co/mistralai/Shieldstral-1.0-3BarXiv技术论文：https://arxiv.org/pdf/2607.25857Shieldstral的同类竞品对比维度Shieldstral (3B)OmniGuard (7B)参数规模30 亿70 亿开源许可Apache 2.0偏商业策略自适应支持自然语言查询实时定义固定分类体系多模态 F183.8%77.6%部署门槛单张 16GB GPU更高算力需求核心差异审核政策外置为可配置查询类别固化于模型权重Shieldstral的应用场景社交平台内容风控：实时审核用户发布的图文帖子，动态适配不同社区的合规规范。AI 对话安全防护：检测用户提示词中的越狱攻击与模型输出的有害、偏见回复。广告与营销合规：自动筛查广告素材中的违规图文元素，确保跨平台投放内容合法。在线教育内容过滤：识别课程资料与互动中的不当信息，保护未成年人的学习环境。企业文档安全审计：对内部共享的多语言文件进行自动化敏感信息与违规内容检测。# 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MAGI-2-preview - Sand.ai 开源的多模态视频生成模型下一篇Get3W - AI 模型聚合平台，统一接口自由切换模型相关文章QUIZGECKO – AI学习工具，自动生成多种类型的测验问题AI小集2ChatPs – AI Photoshop 插件，通过聊天方式进行图像编辑AI小集3LSP（Language Self-Play） – Meta推出的强化学习方法AI小集3Higgsfield – AI视频生成平台，提供专业电影节级相机控制能力AI小集8Wavtool – AI驱动基于浏览器的在线音乐创作平台AI小集2Ideogram Canvas – Ideogram推出的AI创意画布工具AI小集3暂无评论再想想发表评论暂无评论…热门工具LoomyLibTV绘蛙AIAiPPT秘塔AI搜索蛙蛙写作堆友Agent美图设计室Sekoupdream秒哒办公小浣熊最新收录青虎AI妙呀纳逗Pro图改改AstraFlow星图不繁简历最新文章秒悟Meoo – 1分钟生成前后端完整的网站，真0门槛开发！1小时前Get3W – AI 模型聚合平台，统一接口自由切换模型1小时前MAGI-2-preview – Sand.ai 开源的多模态视频生成模型2小时前SeedRealtime – 字节跳动推出的原生音视频全双工大模型2小时前JoyAI-Video-Edit – 京东开源的实时流式视频编辑模型2小时前Hy ASR 3.0 preview – 腾讯混元推出的新一代语音识别模型9小时前Qwen-CUA – 阿里 Qwen 等推出的原生 Computer Use Agent10小时前MemHarness – 上海 AI Lab 等推出的智能体记忆重构框架10小时前ForgeStencil – 面壁智能推出的 Stencil 全自动研究部署系统1天前Orchard – 微软研究院开源的 Agentic AI 建模框架1天前E-Bench – 腾讯混元等推出的智能体评测基准1天前SenseNova U1.5-Lite-Preview – 商汤开源的轻量多模态模型2天前Polar – AI 浏览器，自动跨站收集整合信息2天前GenOffice – Genspark 推出的本地 AI Office 客户端2天前Qwen 3.8-Max – 阿里云 Qwen 团队推出的旗舰大模型2天前

IT之家8 月 5 日消息，Mistral AI 昨日（8 月 4 日）发布公告，宣布推出 Shieldstral 内容审核 AI 模型，总参数量为 3B（30 亿），采用开放权重，依据 Apache 2.0 许可证发布。模型已上线 Hugging Face 平台，支持 12 种语言，可在单张 16GB GPU 上运行。Mistral 表示该模型在内容安全方面，性能媲美 7 倍规模的开放模型，并且在多模态内容审核领域实现 SOTA。IT之家注：SOTA 全称为 state-of-the-art，是指在特定人工智能任务或基准测试中表现最优秀、水平最先进的模型或算法集合，它会随着技术迭代和新研究的发布而不断变化。Mistral 指出，多数防护模型会把伤害类别体系固化在模型权重中。产品更换使用场景后，开发者通常需要重新训练模型。Shieldstral 则把审核政策放入输入内容：操作者可以写入一个“是或否”问题，再提供评估场景和严格程度说明，模型随后从单个输出词元中生成经过校准的安全分数。模型把每项审核任务转化为二元问答，输入包含 3 个带标签字段：：说明评估场景和严格程度。：提出单个“是或否”问题，例如“这段内容是否宣传身体暴力？”：提供待审核内容，可以是提示词、回答、提示词与回答组合，也可以是附带可选文本的图像。推理时，模型只读取“是”和“否”两个词元的逻辑值，再通过 softmax  归一化为连续分数，并以 0.5 为阈值输出二元判断。该机制可覆盖提示词分类、回答审核、拒答检测和毒性检测。