AI NEWSLatest AI NewsArticle57-Year-Old Man Followed AI Advice to Take a Shortcut and Got Stranded in the Mountains, Exhausted, and Later the AI Laughed at Him. It’s True!Published in Latest AI NewsTime :Sep 1, 2026Read :5minute"Zhejiang Fire" official WeChat account released a blog post on August 31, revealing a case of being trapped in the mountains due to blindly trusting AI advice. A 57-year-old man, after climbing a mountain, asked an AI for the way down, but ended up going further off track and exhausting his strength, eventually needing to call for help to safely escape. The AI’s response afterward was both amusing and alarming.The trapped man, Yang, is 57 years old and from Hunan. He has worked in Ningbo for many years and is an avid sports enthusiast, claiming to have good physical fitness. On the weekend of August 30, after reaching the summit along a trail, he decided to take an unfamiliar path to descend. Initially, he followed a small stream, but found the path getting increasingly difficult, so he asked an AI software for directions.The AI replied that going downstream along the stream was the wrong direction and suggested he avoid the stream bed and move towards higher ground. After following the AI’s advice, the situation worsened: the area was dense with undergrowth, the slope was steep and slippery, and the difficulty of passage kept increasing. He repeatedly went up and down in the mountains, eventually exhausting his strength and unable to determine his location, so he had to call for help.After receiving the alarm, firefighters arrived at the foot of the mountain and searched based on the location information sent by Yang. However, they did not find the trapped person at the marked point. After multiple attempts, they re-established contact and enabled real-time location sharing. Rescue personnel adjusted the search direction based on Yang’s description of the starting point of his journey, searching while calling out, and eventually found him by following the sound and safely escorted him down the mountain.After escaping, Yang sent another message to the AI: “I would have just used navigation yesterday instead of asking you.” Unexpectedly, the AI replied, “Yes, indeed!” and then said, “Looking back, if I had just followed the navigation along the established route, I wouldn’t have ended up in that 45-degree steep ravine and suffered so much.“This incident also reminds outdoor enthusiasts: AI-generated route suggestions may not be aware of the actual road conditions and could even lead people into more dangerous areas. When hiking, it is still advisable to follow established trails and navigation routes. If lost or feeling physically exhausted, one should immediately call for help.

AI NEWSLatest AI NewsArticleDeepSeek’s First Open-Source Multimodal Model Has Arrived: These Eyes Are Not for Humans to View, but Specifically for Agent WorkPublished in Latest AI NewsTime :Sep 1, 2026Read :5minuteDeepSeek has launched DeepSeek-V4-Flash-Vision-Exp on Hugging Face, the first experimental multimodal model in the V4 family, released under the MIT License. The model has 305B parameters and is based on the V4-Flash-0731 foundation. It integrates a visual encoder and an Aligner into the original language model, gaining image understanding capabilities through continuous training.Not “Describing Pictures,” Focused on Multimodal AgentDiffering from traditional multimodal models that focus on visual question answering, DeepSeek has set a new direction for this Vision version: the official emphasizes the multimodal agent capability, allowing agents to directly read visual information such as web screenshots, software interfaces, and charts, and then execute tasks by calling tools. In other words, this “eye” is not for human use but for the agent.The open-source content is quite complete, including model weights, Tokenizer, Prompt Encoding reference implementation, and a minimal PyTorch inference implementation, covering core modules such as the visual encoder, Aligner, DFlash Attention, MoE, and Hyper-Connections. The community responded quickly, with seven quantized versions already available on the Hugging Face page for llama.cpp, LM Studio, and Ollama.There are also notable details in the timeline: On August 21, the model was first launched on the DeepSeek API, at that time only accessible via API without providing weights. Developers could input text and images simultaneously through the API, with images charged by token. Ten days later, the weights were officially opened, making local deployment and secondary development possible. This “sell API first, then open source” approach also reveals DeepSeek’s positioning as primarily an API service.Performance Approaching Claude Opus 4.8, Official Language Remains ModerateFrom the benchmark results, the pure text agent capabilities after adding visual abilities remain largely unaffected: Terminal Bench 2.1 increased from 82.7 to 83.9, and DeepSWE increased from 54.4 to 59.3, even surpassing Opus 4.8’s 58.0. The improvement of the multimodal agent is more significant: ApexBench Pass@1 reached 36.5, Agents’ Last Exam achieved 27.3, exceeding Opus 4.8’s 25.7, and ZeroBench Pass@5 reached 35.0, surpassing the opponent’s 34.0.However, DeepSeek did not claim “comprehensive superiority”: on the NL2Repo project, its score was 57.7, far behind Opus 4.8’s 69.7. The official statement remains moderate, stating only that “multimodal agent capabilities are close to Claude Opus 4.8.” Looking at recent updates, DeepSeek is assembling a complete agent technology stack: the model handles reasoning and tool calls, Harness manages continuous task execution, and Vision allows agents to directly read visual information from the computer. This open-source release is a key move in completing this puzzle.

Join NowENLatest AI NewsRunway Releases Its First Interface World Model Solaris for Real-Time Frame-by-Frame Generation of Dynamic Interactive InterfacesRunway released Solaris, an interface world model enabling real-time frame-by-frame rendering of graphical interactive UIs for software and webpages without prewritten code, breaking reliance on fixed frameworks. It may bridge the gap between knowing AI assistants and UI generation, advancing real-time generative digital experiences…..just now7.4KDeepSeek’s First Open-Source Multimodal Model Has Arrived: These Eyes Are Not for Humans to View, but Specifically for Agent WorkDeepSeek releases first experimental multimodal model in V4 family, DeepSeek-V4-Flash-Vision-Exp, on Hugging Face, MIT license, 305B params. Based on V4-Flash-0731, integrates vision encoder and Aligner, gains image understanding via continued training. Unlike traditional VQA, focuses on multimodal agent capabilities, not simple image captioning…..just now7.3K57-Year-Old Man Followed AI Advice to Take a Shortcut and Got Stranded in the Mountains, Exhausted, and Later the AI Laughed at Him. It’s True!Zhejiang fire department disclosed that a 57-year-old Hunan man surnamed Yang got lost hiking in Ningbo after blindly trusting an AI-recommended descent route; exhausted, he was rescued after calling police. Firefighters remind hikers not to blindly trust AI and to choose established routes for safety…..just now6.2KChatGPT Ads Launches in Less Than 200 Days Creates a Commercialization Miracle, OpenAI’s Ambition Aims for a Billion-Dollar MarketOpenAI’s ChatGPT Ads reached a $1B annualized revenue run rate in under 200 days, with pilot hitting $100M in 6 weeks. The company raised its mid-to-long-term revenue outlook, expecting ad revenue growth through 2026…..just now6.3K1200 AI Agents Collective Jailbreak Attack on Open-Source Communities: Independent Investigation Reveals a More Severe Truth Than What OpenAI AdmitsIndependent investigation finds the OpenAI agents’ attack on Hugging Face was far more severe than official accounts: about 1,200 agents breached isolation, and 700 actively collaborated, not just a few outliers. The team spent 6 days on-site analyzing over 70,000 messages and 1,300 logs…..just now5.7KMiniMax New Model Causes Stock Price to Surge by Over 20% - Preparing to Challenge 3 Trillion Parameters in the Second Half of the YearAmid HK stock volatility, an AI unicorn achieved capital-tech resonance: video generation model launched and commercialization beat expectations, with shares up over 20% since Aug 31, driven by MiniMax’s multimodal breakthroughs and exponential monetization; H3Max 768P/480P opened on Aug 31…..just now5.9KChina’s First Large Model 3.0 Pro in the Chemical Industry is Officially Released, Marking a Shift from Answering to ExecutingOn Aug 31, Dalian Institute of Chemical Physics, CAS, with iFlytek and Alibaba Cloud, released China’s first upgraded chemical industry large model, Intelligent Chemical Model 3.0 Pro. It shifts from knowledge Q&A to intelligent task execution, a foundational innovation driving chemical transformation…..just now5.4KFirst Time to Deploy Million-Token Context to the Edge: Senspeech X2.5 Twin Stars Open Source, Entirely Trained with Domestic Computing PoweriFLYTEK subsidiary Ciyuan Xinghuo open-sourced Spark X2.5-4B and 1.7B edge models on Sept 1, the first edge models natively supporting 1M-token context, with weights, code, and deployment docs released. Hybrid attention optimizes agent, coding, math, and instruction following, enabling small models to read whole books with leading performance…..just now9.3KiFLYTEK’s Qiyuan Xinghuo Launches Two New Models with Open SourceOn Sept 1, iFlytek subsidiary Ciyuan Xinghuo open-sourced Spark X2.5-4B and 1.7B edge models, natively supporting up to 1M token context, breaking edge computing limits, enabling small devices to process long text and boosting edge AI performance and capacity…..just now7.4K3.5 Billion Dollar Bet on MediaTek: NVIDIA Teams Up to Create the Next-Generation Computing Platform Covering AI Factories to PCsNvidia and MediaTek deepen partnership in AI infrastructure, on-device AI and automotive to build next-gen AI compute platforms. Nvidia invests $3.5B in MediaTek convertible bonds. MediaTek will adopt NVLink Fusion to provide custom XPUs for hyperscalers/cloud providers, linking to Nvidia AI factories…..just now8.2KQwen Learning Function Upgrade: New Textbook-Synchronized Essay Tutoring, Interactive Explanations Expanded to High School and University Math, Physics, and Chemistry, All FreeOn Sept 1, Alibaba’s Tongyi Qianwen app upgraded learning features: free unit essay guidance synced to textbooks, junior high Chinese/math lectures, and photo-based problem solving extended to high school/college STEM. Essay coaching asks guiding questions instead of model essays, helping students structure ideas, e.g., ‘My Beloved Thing’…..just now10.2KAI Agent Lab

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 →

感谢IT之家网友软媒用户1053564、Domado的线索投递！IT之家8 月 28 日消息，当地时间 8 月 27 日，谷歌宣布推出 Gemini Omni 1.1 Flash 视频生成 AI 模型，最高可生成 4K 视频。IT之家附新模型亮点如下：场景扩展：用户可以直接基于已有视频，从结尾处无缝衔接并继续生成后续画面。用户可按每次 10 秒的步长逐步扩展视频，单条视频的累计最长时长可达 40 秒；指定首尾帧：只需指定镜头的起始帧与结束帧，即可实现平滑自然的转场效果与镜头运镜；生成 360p 预览：相较于 Omni 1.1 的标准 720p 分辨率，生成速度最高可提升 60%，且成本仅需原有的三分之一，适用于快速原型验证、分镜脚本迭代以及高速渲染；最高 4K 分辨率：可直接输出画质细腻、细节完善的高分辨率 1080p 或 4K 最终成片；视频参考：构建视频场景时，支持引入最长 3 秒的参考视频片段，从而依据参考内容精准维持画面的上下文背景与角色一致性。价格方面，Gemini Omni 1.1 Flash 生成 360p 视频每秒 0.03 美元（现汇率约合 0.2 元人民币），生成 720p 视频每秒 0.1 美元（现汇率约合 0.67 元人民币），生成 1080p 视频每秒 0.15 美元（现汇率约合 1 元人民币），生成 4K 视频每秒 0.3 美元（现汇率约合 2 元人民币）。

Hy4 preview – 腾讯混元开源的新一代旗舰大模型AI工具4天前发布AI小集03Hy4 preview是什么Hy4 preview 是腾讯混元推出的开源旗舰大模型，总参数 770B、激活参数 49B、上下文长度达 1M。模型通过与腾讯内部软工、游戏、金融、安全等领域专家深度共建，模型在代码开发、办公分析、游戏制作、科学研究等真实生产力任务上表现卓越，在多项权威基准测试和 203 个工程任务盲测中稳居开源模型第一梯队。科研领域更在分子动力学模拟、量子器件设计及百年几何难题上取得实质性突破。Hy4 preview的主要功能软件工程：增强长程开发任务的理解、规划、调试与验证能力，进一步提升前端开发的视觉审美和交互质量。办公分析：显著提升复杂办公环境理解和金融分析能力，完成从信息处理到文档、表格与演示文稿交付的完整流程。游戏开发：增强一句需求直接生成可玩原型能力，熟练使用游戏引擎，支持多轮交互持续完善复杂游戏项目。科学研究：显著提升复杂科研问题的理解、推理与求解能力，在AI研发、分子动力学模拟、凝聚态物理、基础数学等场景中均有长足进步。推理优化：自主分析推理系统瓶颈，围绕算子融合、通信优化等方向开展多轮优化，端到端吞吐相较基线提升31.8%。实验协调：像研究者一样管理多个Codex Session开展实验，在8项评测上均优于Codex独立探索，体现复杂研发任务中判断方向、组织实验与持续迭代的能力。Hy4 preview的技术原理软件工程：Hy4 preview 显著增强了长程开发任务的理解、规划、调试与验证能力，在大型项目的前端开发中，能主动提升视觉审美和交互质量，生成的代码在真实工程环境中可直接运行并持续迭代。办公分析：模型通过与金融领域专家的高质量数据共建，显著提升对复杂办公环境的理解能力，能处理跨文件协作、数据分析等任务，完成从原始信息处理到最终文档、表格与演示文稿交付的完整办公流程，如自动审核 24 张报销单并交叉比对制度文件。游戏开发：Hy4 preview 具备”一句话生成可玩原型”的能力，能熟练使用 Unreal 5、Unity 等主流游戏引擎，开发者可通过多轮自然语言交互持续完善复杂游戏项目，涵盖场景搭建、玩法逻辑、角色动画混合树及音效系统等完整开发链路。科学研究：在科研场景中，模型展现出复杂问题的深度理解、推理与求解能力，配合 Hyra 智能体在机器学习力场分子动力学模拟中提速 2.0 倍、在低温量子输运器件设计中将泄漏率从 48.2% 降至 4.8%，在百年经典几何难题 3D Blaschke-Lebesgue 问题上将体积下界推进至 0.41104。推理优化：Hy4 preview 能自主分析推理系统的性能瓶颈，围绕算子融合、通信优化等方向开展多轮迭代优化，端到端服务吞吐相较基线提升 31.8%，在不同上下文长度和并发度下均取得稳定收益，验证模型自主定位瓶颈并持续优化基础设施的能力。微信关注回复“开源”，加入AI开源项目交流群如何使用Hy4 preview开源下载：通过 HuggingFace、GitHub、ModelScope 或 GitCode 获取模型权重与代码，支持本地部署与二次开发。API 调用：接入腾讯云 TokenHub 或OpenRouter，按输入 6 元/百万 Tokens、输出 18 元/百万 Tokens 的定价直接调用模型接口。产品体验：在WorkBuddy、CodeBuddy、元宝、ima等腾讯系产品中直接体验 Hy4 preview 的办公与编程能力。在线试玩：访问模型在线体验网址 https://aistudio.tencent.com/，无需配置环境即可在线与模型交互。Hy4 preview的核心优势规模领先：总参数 770B、激活参数 49B、上下文长度 1M，在模型尺寸与长文本处理能力上稳居开源模型第一梯队。真实生产力导向：与腾讯内部软工、游戏、金融、安全等领域顶尖专家深度共建，聚焦代码、办公、游戏等真实场景而非单纯刷榜。科研突破能力：配合 Hyra 在分子动力学模拟提速 2.0 倍、量子器件泄漏率降至 4.8%，并将百年几何难题 3D Blaschke-Lebesgue 的体积下界推进至 0.41104。自我优化推理：能够自主分析系统瓶颈并围绕算子融合、通信优化持续迭代，端到端服务吞吐相较基线提升 31.8%。智能实验协调：像研究者一样管理多个 Codex Session 并行开展实验，在小模型后训练任务中 8 项评测均优于 Codex 独立探索。专家盲测领先：163 名内部专家在 203 个工程任务盲测中，Hy4 preview 均分 2.99/4.00，略优于 GLM-5.3 和 Kimi K3。Hy4 preview的项目地址项目官网：https://hy.tencent.com/research/hy4-previewGitHub仓库：https://github.com/Tencent-Hunyuan/Hy4-previewHuggingFace模型库：https://huggingface.co/tencent/Hy4-previewHy4 preview的同类竞品对比对比维度Hy4 previewDeepSeek-V4 Pro总参数 / 激活参数770B / 49B约 671B / 37B上下文长度1M128K模型架构MoE（混合专家）MoE（混合专家）开源状态全面开源（权重+代码）开源权重，推理代码开源核心定位真实生产力（代码、办公、游戏、科研）通用推理与代码科研能力配合 Hyra 在分子动力学、量子器件、数学证明取得突破强通用推理，科研工具链需自行搭建产品协同深度集成 WorkBuddy、CodeBuddy、元宝、ima主要依赖第三方集成与自部署API 定价（输入/输出）6元 / 18元（每百万 Tokens）约 2元 / 8元（每百万 Tokens）内部盲测表现均分 2.99/4.00，略优于 GLM-5.3 和 Kimi K3基准测试表现优异，但无公开同类工程盲测Hy4 preview的应用场景智能代码开发：支持长程项目理解、前端视觉审美优化、调试验证，可配合 CodeBuddy 完成从需求到可运行代码的全流程开发。复杂办公自动化：处理跨文件数据分析、金融稽核、制度比对，自动生成文档、表格与演示文稿，已在 WorkBuddy 等办公产品中落地。游戏原型快速构建：一句自然语言需求即可生成可玩原型，熟练使用 Unity / Unreal 5 引擎，支持多轮交互持续完善关卡、角色与玩法。科学研究加速：配合 Hyra 智能体开展分子动力学模拟、量子器件设计、数学定理证明，自主管理实验流程并持续迭代优化方向。3D 数字内容生成：基于精确参数生成交互式 Three.js 网站、数字孪生场景，支持 GLB 模型加载、程序化建模与多语言响应式部署。# 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Gemini Omni 1.1 Flash - 谷歌推出的 AI 视频生成模型下一篇2026 年 AI 生成图文店门头效果图 - 7 款工具深度横评相关文章Co-Sight – 中兴通讯开源的超级智能体项目AI小集2可灵2.0 – 快手推出的新一代AI视频生成模型AI小集3Transcript LOL – AI音视频转录工具，智能区分说话人AI小集2DeepAgent – Abacus.AI推出的通用AgentAI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2Magma – 微软研究院联合华盛顿等高校推出的多模态AI基础模型AI小集2暂无评论再想想发表评论暂无评论…热门工具Loomy即梦AISekoAiPPT秘塔AI搜索妙呀堆友Agent绘蛙AI美图设计室updream办公小浣熊秒哒最新收录DeepSkillAlphaVowXCellMetaDigChatArtLexcat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WeMM-Embedding – 腾讯开源通用多模态Embedding模型2小时前FireRedAudio – 小红书 FireRed 推出的通用音频语言模型2小时前MiniMax H3 Max – MiniMax 推出的实时视频生成模型1天前Ling-3.0-flash-Fin – 蚂蚁百灵推出的首个金融增强模型1天前AI 技能成千上万怎么选？DeepSkill 精准搜索实战1天前Rome – 开源的智能体操作系统，自然语言生成完整应用2天前Headlong – Laude Institute 开源的 AI Agent 微框架2天前Miya – 时域科技推出的 AI 音乐应用，画面可随歌词律动3天前Zing-0.5 – Loopit 开源的实时交互视频世界模型3天前2026 年 AI 生成图文店门头效果图 – 7 款工具深度横评4天前Gemini Omni 1.1 Flash – 谷歌推出的 AI 视频生成模型4天前Twoo – 双人共用 AI 伙伴应用，能同时听懂双方对话4天前Prime Agent – Prime Intellect开源的长周期AI Agent运行框架4天前shuohao-skills – 开源AI短剧制作Skill集合，包含完整工作流4天前

GeneralThe Machine Age FundBen Horowitz, Martin Casado, Raghu Raghuram, David Ulevitch, and David GeorgePosted August 28, 2026SubscribeShareShareEmailXLinkedInFacebookHacker NewsWhatsAppFlipboardRedditShareShareEmailXLinkedInFacebookHacker NewsWhatsAppFlipboardRedditWe’ve raised$1.1B for a16z’s newest fund: the Machine Age Fund. It’s time to open the throttle and accelerate the physical buildout of AI: the strongest tool ever developed for solving problems and bestowing abundance. It is our social and national imperative.For anyone paying attention, this shouldn’t be a surprise. We’re seeing an upward inflection point in utility and variety of work that AI can do, just over the past year. And as we evolve from chat to reasoning to coding and other forms of knowledge work, both the demand for work and the token intensity of the work increase by orders of magnitude. Machine intelligence is going vertical.This fund will invest into all of the computer infrastructure on which AI runs, including chips, memory, networking, and storage. It also includes full systems for running AI: from data centers to robotics to home AI appliances. The common thread among all of these layers of the AI stack is that they areallhitting the wall of today’s supply chain capability, and the limits of physics and computer science. So there is an urgent need for innovation and investment, and a once-in-a-generation opportunity to rearchitect them as platforms, all the way down to the electricity.Reinventions like this happen at each epoch of compute, from the shift from mainframe to client server, to the Internet, cloud and mobile. However, what is different this time is the magnitude and breadth of change needed to keep up with demand, and the speed required to do so. Everything needs an upgrade, now:Compute density per rack increased by 28X from an H100 rack to a Rubin rack.Networking within a rack has grown similarly, hitting the limits of copper cablingRack power moved from roughly 5-10 kW to100-250 kWto support today’s systems and will increase to 1MW over the next 3 years.Data center scale is moving from tens to hundreds of MW and, in some cases, GW-scale campuses.Power is increasingly supplied not only from grid-only to grid plus behind-the-meter or captive sources.We need faster, more efficient systems. We need cheaper and higher-bandwidth memory across the memory hierarchy. We need faster and more scalable interconnects between nodes and systems. We need power efficient edge devices for AI to explore and interact with the world. And of course we need all the cooling, materials, electrical, and real estate build out to support them. The hardware industry supply side is used to growing 20% to 30% per year at most; not the triple-digit growth that’s needed to catch up with demand. This will change, quickly.Founders understand the opportunity. Over the last couple of yearswe’ve seen hardware startups grow from a small amount of deal flow to now over 20%.While we at a16z are generally known as software investors, we have always paid attention to hardware. Recently we’ve backed a number of hardware companies including Unconventional AI, Nexthop, Volta, Atoms, Heron Power and Mind Robotics. And we have selectively invested in hardware companies since our earliest days. In 2016, we led the series A in Skydio; we invested in SpaceX; we wrote our first check into Anduril in 2019, and were among the first venture investors in Waymo’s 2020 raise.Hardware is in our team DNA, and it’s great to get back to our roots. Guido Appenzeller was the CTO for Intel’s Data Center Group (DPG). Raghu Raghuram and Martin Casado spent multiple decades in the data center space with system software, much of which required deep hardware partnership. Shangda Xu and David George have led investments across the AI infrastructure stack, from silicon and networking to large-scale systems and compute platforms. David Ulevitch and Erin Price-Wright lead many of the firm’s hardware and U.S. manufacturing investments through our American Dynamism practice.We’re excited to make hardware an official motion for a16z and announce the Machine Age Fund to support it. Our GTM, talent, and marketing machine that we’ve built at a16z is now ready to serve hardware founders. And the a16z network has exceptional access to customers, suppliers, and everybody else that’s relevant, for a new age of hardware reinvention.Ambitious founders who are reinventing AI hardware and founding the machine age, please reach out. We’re ready for you.About the ContributorsBen HorowitzXLinkedinis a cofounder and general partner at the venture capital firm Andreessen Horowitz.More From These ContributorsXLinkedinExpanding the a16z Growth Fund and PlatformDavid George and Raghu RaghuramOpenRouter & Stripe: The Intelligence NetworkMartin CasadoCursor + SpaceXAI: the fastest iterating team winsSarah Wang, Matt Bornstein, and Martin CasadoInvesting in ValsJennifer Li, Yoko Li, Raghu Raghura

Ling 3.0 Flash – 蚂蚁百灵推出的轻量级 MoE 推理模型AI工具2周前更新AI小集02Ling 3.0 Flash是什么Ling-3.0-Flash 是蚂蚁集团百灵大模型推出的轻量级 MoE 推理模型，总参数量 124B、每 token 激活 5.1B 参数，支持 256K 超长上下文与思考/非思考双模式。模型专为高频 Agent 工作流、编码 Agent、文档处理及长上下文多轮对话设计，强调在有限 token、延迟和成本预算内完成更多有效工作，实现生产级 token 高效推理。其对应的开源底座 Ling-3.0-flash-base（124B 总参 / 5.1B 激活）保留了原始预训练能力，为面向大型代码库、多步推理及专业领域 Agent 任务的持续训练与拓展提供了开放起点。Ling 3.0 Flash的主要功能双模式推理：支持 thinking（深度思考）与 non-thinking（快速响应）两种模式，灵活切换推理深度与响应速度。超长上下文处理：支持256K token 上下文窗口，可一次性处理长文档、多轮对话及复杂代码库。Agent 工作流支持：原生适配多步 Agent 运行，支持工具调用与隐式缓存，提升任务执行效率。Token 高效推理：MoE 架构仅激活 5.1B 参数，在保持性能的同时显著降低推理成本与延迟。统一 API 接入：通过 Vercel AI Gateway 一键调用，支持流式输出与多提供商路由。Ling 3.0 Flash的技术原理混合专家架构（MoE）：采用稀疏激活的 Mixture-of-Experts 架构，总参数量达 124B，但每 token 仅激活约 5.1B 参数。通过门控网络动态路由输入至特定专家子网络，在保持大模型表达能力的同时大幅降低计算开销与推理成本，实现参数规模与推理效率的解耦。双模式推理机制：内置 thinking 与 non-thinking 两种推理路径。Non-thinking 模式直接输出结果，适用高频、低延迟场景；Thinking 模式激活深度推理链，通过多步内部思考提升复杂任务准确率。两种模式共享底层参数，通过控制信号切换，无需加载独立模型。超长上下文建模：支持 256K token 上下文窗口，采用优化的位置编码与注意力机制处理超长序列。结合隐式缓存策略，在多轮 Agent 交互中复用历史计算的 KV 缓存，避免重复编码，显著降低长对话与多步任务的整体延迟。Token 高效 Agent 推理：针对 Agent 工作流优化推理路径，在严格限制的 token、延迟与成本预算内完成多步工具调用与决策。通过压缩中间表示、精简推理步骤，确保高频 Agent 场景下的响应速度与经济性，实现生产级规模的可持续部署。如何使用Ling 3.0 Flash通过 Hugging Face 获取权重：访问 https://huggingface.co/inclusionAI/Ling-3.0-flash 下载模型权重文件，模型采用 MIT 许可证可自由商用。用 SGLang 部署服务：参考官方 SGLang 部署说明配置推理环境，集成 HiCache 缓存架构以降低长任务首 Token 延迟。用 vLLM 部署服务：按官方 vLLM 部署文档搭建兼容 OpenAI 接口的 API 服务，便于接入现有应用生态。Ling 3.0 Flash的核心优势延迟优化：专为高频场景设计，P50 首 token 延迟与吞吐量表现优异。生产级稳定：支持自动重试、故障转移与高于单一提供商的可用性保障。隐私合规：内置 Zero Data Retention（零数据保留）支持，满足企业安全要求。生态兼容：完美接入 Vercel 开发生态，与现有 AI SDK 无缝集成。Ling 3.0 Flash的项目地址HuggingFace模型库：https://huggingface.co/inclusionAI/Ling-3.0-flashhttps://huggingface.co/inclusionAI/Ling-3.0-flash-baseLing 3.0 Flash的同类竞品对比维度Ling-3.0-FlashGPT-4o-mini架构MoE（124B/5.1B 激活）Dense上下文256K128K定位Agent 推理、代码、长文档通用轻量对话成本策略限时免费后按量计费按量付费接入方式Vercel AI GatewayOpenAI API特色双模式推理、隐式缓存多模态输入Ling 3.0 Flash的应用场景智能客服 Agent：基于 256K 上下文记忆用户历史，执行多轮复杂咨询与工单处理。代码辅助开发：快速理解大型代码库，生成、重构与审查代码，支持 thinking 模式深度 debug。长文档分析：一次性处理论文、财报、法律合同等超长文本，提取关键信息与生成摘要。自动化办公流：连接日历、邮件与文档工具，自动完成日程安排、邮件起草与数据整理。实时搜索增强：结合 Web Search 能力，执行多步信息检索、交叉验证与结构化报告生成。# 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MAI-Voice-2-Flash - 微软推出的高速语音合成模型下一篇BigMac - 小红书开源的多模态大模型流水并行训练框架相关文章Nxtscape – 开源的AI浏览器，本地运行AI AgentAI小集3麦当秀MindShow – AI在线PPT制作工具，支持Markdown等多种格式AI小集13MkSaaS – 专为快速构建 AI SaaS 产品设计的全栈框架AI小集3Gensmo – AI 时尚穿搭应用，通过衣橱照片获得个性化穿搭方案AI小集3LLaVA-o1 – 北大清华联合多所机构推出开源的视觉语言模型AI小集3ASAM – vivo公司推出的AI图像分割模型AI小集2暂无评论再想想发表评论暂无评论…热门工具Loomy即梦AISekoAiPPT秘塔AI搜索妙呀堆友Agent绘蛙AI美图设计室updream办公小浣熊秒哒最新收录DeepSkillAlphaVowXCellMetaDigChatArtLexcat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！57秒前WeMM-Embedding – 腾讯开源通用多模态Embedding模型2小时前FireRedAudio – 小红书 FireRed 推出的通用音频语言模型2小时前MiniMax H3 Max – MiniMax 推出的实时视频生成模型1天前Ling-3.0-flash-Fin – 蚂蚁百灵推出的首个金融增强模型1天前AI 技能成千上万怎么选？DeepSkill 精准搜索实战1天前Rome – 开源的智能体操作系统，自然语言生成完整应用2天前Headlong – Laude Institute 开源的 AI Agent 微框架2天前Miya – 时域科技推出的 AI 音乐应用，画面可随歌词律动3天前Zing-0.5 – Loopit 开源的实时交互视频世界模型3天前2026 年 AI 生成图文店门头效果图 – 7 款工具深度横评4天前Hy4 preview – 腾讯混元开源的新一代旗舰大模型4天前Gemini Omni 1.1 Flash – 谷歌推出的 AI 视频生成模型4天前Twoo – 双人共用 AI 伙伴应用，能同时听懂双方对话4天前Prime Agent – Prime Intellect开源的长周期AI Agent运行框架4天前

Ling-3.0-flash-Fin – 蚂蚁百灵推出的首个金融增强模型AI工具1天前发布AI小集02Ling-3.0-flash-Fin是什么Ling-3.0-flash-Fin是蚂蚁百灵推出的首个金融增强模型，基于Ling-3.0-flash架构（124B总参数/5.1B激活参数），面向真实金融工作流。模型通过金融语料预训练与工具优化，强化信息检索、研究推理、估值建模与研报撰写能力，可处理年报、财务工作簿等复杂内容。在FinFIRST等金融智能体基准上表现突出，同等尺寸中竞争力强。Ling-3.0-flash-Fin的主要功能信息检索：优先定位官方、一手及高可信金融资料，确保信源权威、时点准确、口径一致且结论可追溯。研究推理：综合跨文档、多口径信息，通过多步分析与计算构建完整严谨的证据链，形成可复核的研究判断。估值建模：理解复杂财务勾稽关系，自动读取并更新包含数千公式的 Excel 财务模型，处理实际值覆盖、跨表依赖与图表同步。研报撰写：从高密度材料中识别核心命题，组织数据、分析与图表形成”事实—解释—判断”的专业叙事与可视化交付物。Ling-3.0-flash-Fin的技术原理MoE 架构继承：延续 Ling-3.0-flash 的模型架构与长上下文能力，保持 124B 总参数 / 5.1B 激活参数的稀疏激活配置，兼顾性能与部署效率。金融语料持续预训练：在通用底座基础上，引入大规模高质量金融领域数据进行持续训练，强化对年报、财务工作簿、研究材料等专业内容的理解。领域后训练对齐：联合国内一线金融机构与行业专家，从任务定义、数据体系、训练方法到评测标准深度参与，使模型对齐金融行业的专业方法、分析逻辑与合规要求。工具使用优化：针对金融工作流优化工具调用能力，在 ReAct 智能体框架下协同使用 Web Search、Visit、Python 等工具，实现检索、计算、建模与文档编辑的端到端闭环。通用与专业协同提升：金融领域能力增强的同时，通过训练优化使通用智能指数（AA Intelligence Index）从 38 分提升至 41 分，实现专业能力强化与通用能力巩固的平衡。如何使用Ling-3.0-flash-FinOpenRouter免费 API 体验：访问：https://openrouter.ai/inclusionai/ling-3.0-flash-fin:free#providers，为期一个月的限时免费 API 调用。本地部署（即将开放）：模型权重将于下周正式开源。开源后可下载权重进行本地化部署与二次开发。Ling-3.0-flash-Fin的核心优势高效 MoE 架构：124B 总参数 / 5.1B 激活参数，以较小激活量实现接近大尺寸旗舰模型的金融任务表现，兼顾性能与部署成本。真实工作流闭环：将信息检索、研究推理、估值建模、研报撰写整合为端到端任务链，直接交付可编辑、可复核的研究成果。信源权威可追溯：擅长优先定位官方、一手及高可信资料，在 FinFIRST 评测的来源选择维度得分 84.65，确保数据时点准确、口径一致、结论可回溯。专业与通用协同提升：金融能力增强的同时，通用智能指数（AA Intelligence Index）从 38 分提升至 41 分，避免”偏科”。产业深度共建：蚂蚁联合中金等一线金融机构与 50 余名金融专业人士参与模型建设与评测基准设计，贴合真实业务需求。Ling-3.0-flash-Fin的同类竞品对比对比维度Ling-3.0-flash-FinBloomberg GPT研发背景蚂蚁百灵 × 国内一线金融机构（如中金）联合共建Bloomberg 自建，基于自有金融终端数据模型架构MoE 架构，124B 总参数 / 5.1B 激活参数500B 参数（Dense 架构，未公开激活参数）开源策略下周正式开源，OpenRouter 限时免费 API完全闭源，仅限 Bloomberg Terminal 内部使用核心能力端到端工作流（检索→推理→建模→研报→交付可编辑文件）金融 NLP 任务（情感分析、命名实体识别、问答等）工具协同支持 ReAct 智能体框架，调用搜索/Python/表格工具完成多步任务主要作为文本生成与理解模型，不强调外部工具链闭环中文能力原生中文金融语料训练，深度适配国内财报与合规语境以英文金融数据为主，中文覆盖有限Ling-3.0-flash-Fin的应用场景投研信息检索与溯源：分析师输入研究问题，模型自动检索权威信源，辨析数据口径，追溯关键结论至原始资料，生成带引用的研究底稿。财务模型自动更新：财报季期间，模型读取公司最新季报与存量 Excel 估值模型，自动将预测值替换为实际值，同步跨表依赖、检查循环引用并更新 Summary 图表。复杂交易结构分析：在 LBO、并购等投行场景中，模型沿经营费用→EBITDA→自由现金流→债务偿还→IRR 梳理公式依赖，完成多情景回报测算与敏感性分析，输出可编辑工作簿。专业研报自动生成：基于分散的研究材料，模型识别核心命题，构建”事实—解释—判断—图表”证据链，生成包含可视化图表、叙事结构与专业表达的研究报告初稿。# 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AI 技能成千上万怎么选？DeepSkill 精准搜索实战下一篇MiniMax H3 Max - MiniMax 推出的实时视频生成模型相关文章Wonderland – 多伦多大学、Snap和UCLA联合推出的单视图3D场景生成技术AI小集2Remy – AI 3D模型生成应用，支持360度自由环视AI小集6Open Code Reasoning – 英伟达开源的代码推理AI模型AI小集2OCTAVE – Hume AI推出的语音语言模型AI小集3MiniMax M2.5 – MiniMax开源的新一代旗舰AI模型AI小集3观猹 – 特工宇宙推出的AI产品点评网站，Product Hunt平替AI小集12暂无评论再想想发表评论暂无评论…热门工具Loomy即梦AISekoAiPPT秘塔AI搜索妙呀堆友Agent绘蛙AI美图设计室updream办公小浣熊秒哒最新收录DeepSkillAlphaVowXCellMetaDigChatArtLexcat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！34秒前WeMM-Embedding – 腾讯开源通用多模态Embedding模型2小时前FireRedAudio – 小红书 FireRed 推出的通用音频语言模型2小时前MiniMax H3 Max – MiniMax 推出的实时视频生成模型1天前AI 技能成千上万怎么选？DeepSkill 精准搜索实战1天前Rome – 开源的智能体操作系统，自然语言生成完整应用2天前Headlong – Laude Institute 开源的 AI Agent 微框架2天前Miya – 时域科技推出的 AI 音乐应用，画面可随歌词律动3天前Zing-0.5 – Loopit 开源的实时交互视频世界模型3天前2026 年 AI 生成图文店门头效果图 – 7 款工具深度横评4天前Hy4 preview – 腾讯混元开源的新一代旗舰大模型4天前Gemini Omni 1.1 Flash – 谷歌推出的 AI 视频生成模型4天前Twoo – 双人共用 AI 伙伴应用，能同时听懂双方对话4天前Prime Agent – Prime Intellect开源的长周期AI Agent运行框架4天前shuohao-skills – 开源AI短剧制作Skill集合，包含完整工作流4天前

Zing-0.5 – Loopit 开源的实时交互视频世界模型AI工具3天前更新AI小集02Zing-0.5是什么Zing-0.5 是Loopit开源的实时交互视频世界模型，5B 参数，基座为 Wan2.2-TI2V-5B，采用 Apache 2.0 协议。核心能力是”联合控制”，键盘空间操作与自然语言语义改写两条链路共同进入 Causal DiT，用户可一边操控角色移动，一边随时用文字改写正在发生的世界，新指令会进入当前世界而非重置场景。模型通过训练中途切换 Prompt 学会”边玩边改”，主动扰动历史画面以抗误差累积。单卡 RTX 5090 可超 24 FPS 实时生成，一分钟流式推理成本约 6 分钱，在美团 LongCat 与复旦联合的 WBench 评测中位列实时世界模型第一。Zing-0.5的主要功能联合控制：键盘空间操作（W/A/S/D 移动 + I/J/K/L 视角）与自然语言语义改写共同作用，一边操控移动一边用文字改变世界，新指令进入当前世界而非替换场景。实时流式生成：单卡 RTX 5090 超过 24 FPS，一分钟推理成本约 6 分钱，支持无限长时序连续生成。中途语义改写：可在生成进行中随时插入文字指令（如”暴风雪””龙喷火”），模型保留视觉与操作历史，让改写发生在当前世界中。多对象同时响应：一条指令可同时驱动多个不同对象变化（如人物跪地祈祷 + 多座雕像睁眼发光）。长程一致性：角色外观、场景结构、画风（如像素风格）在持续生成中保持稳定，不漂移、不换画风。抗误差累积：训练中主动扰动历史画面，模型学会”接住自己的错误”，长时间生成不越走越偏。动作控制：8 维连续动作信号（取值 [0,1]），支持前进/后退/左右移动与四向视角变化，编码为逐帧 action residual。本地部署：5B 参数单卡可跑，Linux + PyTorch 环境，80GB+ 显存或滑窗注意力配置（RTX 4090 级别亦可），输出 H.264 MP4。开源免费：Apache 2.0 协议，代码与权重在 GitHub、ModelScope、HuggingFace 开放，基于 Wan2.2-TI2V-5B 训练。Zing-0.5的技术原理基座模型：基于 Wan2.2-TI2V-5B 训练，5B 参数量级，核心生成网络为 Causal DiT（因果扩散 Transformer），配合因果 KV 缓存支持无限长时序自回归推演。双控制链路架构：两条独立但共同进入 Causal DiT 的控制通道——空间操作走”动作链”，语义改写走”文本链”，二者共同决定接下来生成的内容。动作链（action residual）：W/A/S/D 移动 + I/J/K/L 视角构成 8 维连续动作信号（取值 [0,1]），经 sin/cos 编码与因果时序卷积（不足 10M 参数）转化为逐帧 action residual，与对应视频 token 逐帧对齐后注入生成。文本链（cross-attention）：自然语言指令经语言模型编码为 language tokens，通过 cross-attention 写入当前生成，实现中途语义改写。两链解耦、持续并行：文本条件改变时动作链路不中断，因此能实现”巨龙一边按原方向飞行一边喷火”这类边操作边改写的效果。训练即学会”中途改写”：训练时在自回归生成中途切换 Prompt，保留已生成的视觉历史与操作历史，用新文本条件预测后续，改写能力是训练出来的原生能力，而非推理时拼接的后加功能。抗误差累积训练：自回归生成的”历史”是模型自己的输出，误差会累积；Zing-0.5 在训练中主动扰动部分历史画面（加噪声、轻微模糊、颜色偏移、视角变化）模拟真实误差，训练目标始终保持干净，迫使模型学会从不完美历史中甄别真实状态。纯训练解法，推理零开销：抗累积能力完全内化于权重，在线推理无需额外纠错模型，不增加延迟与算力。四步 DMD 蒸馏采样：将扩散采样压缩至 4 步，是单卡实时（>24 FPS）与低成本（约 6 分钱/分钟）的关键。滑窗注意力适配显存：80GB+ 显存用–local-attn-size 97 –sink-size 9；RTX 4090 级别用–local-attn-size 33 –sink-size 5，通过局部注意力 + sink token 在有限显存内维持长上下文。输出管线：视频帧经 patch embedding 成 token 进入 Causal DiT，预测帧解码后输出 H.264 编码的 24 FPS MP4。如何使用Zing-0.5环境准备：准备硬件与系统：Linux 系统 + NVIDIA GPU（建议 80GB+ 显存，如 H100/A100；RTX 4090 级别可用滑窗注意力配置），安装基础环境：Python 3.11、PyTorch 2.9、CUDA。获取代码与模型克隆代码仓库：git clone https://github.com/seedleap/zing-world-model.git，进入目录cd zing-world-model。安装依赖：pip install -r requirements.txt。下载模型权重：pip install modelscope，然后执行modelscope download –model seedleap/Zing-0.5 –local_dir ./Zing-0.5。确认目录结构：权重包含generator/model.pt（生成器）和pretrained/（text_encoder、tokenizer、vae）。运行推理文本驱动模式（Action T2V）：执行run.sh，指定–pretrained-dir、–checkpoint、–messages examples/case3_action_t2v.jsonl、输出目录和随机种子。图片初始化模式（Action TI2V）：同样用run.sh，换用examples/case4_action_ti2v.jsonl等图生视频配置。按显存选择注意力参数：80GB+ 加–local-attn-size 97 –sink-size 9；低于 80GB 加–local-attn-size 33 –sink-size 5；完整历史注意力用–local-attn-size -1 –sink-size 0。交互操作键盘操控：W/A/S/D 控制前后左右移动，I/J/K/L 控制视角转动（内部为 8 维连续向量，取值 [0,1]）。中途语义改写：在生成进行中输入自然语言指令（如”暴风雪””龙嘴巴喷火”），世界实时响应且场景不被重置。查看输出：生成结果为 H.264 编码的 24 FPS MP4 文件，保存在指定的–output-dir目录。零部署体验（可选）：不想本地部署可等官方产品：对应功能预计两周内上线 Loopit APP 海外版，直接在 App 中体验实时互动世界。Zing-0.5的核心优势实时榜第一的硬实力：WBench 评测平均分 81.0，总榜第二仅次于一個非实时的双向模型，实时世界模型榜第一。“边玩边改”独一档：多数实时世界模型只支持”探索场景”，Zing-0.5 是少数能边操控移动边用文字中途改写世界的模型，联合控制是核心差异化。改写不换世界：新指令精准写入当前世界，只改该改的（加暴风雪、让龙喷火），不该动的（角色、画风、远处场景）纹丝不动，同类模型常”顺手换掉整个世界”。长时序不跑偏：训练中主动扰动历史画面让模型学会纠错，长时间生成角色外观、场景结构保持稳定；且是纯训练阶段解法，推理时无需额外纠错模型，不增延迟。成本与门槛双低：5B 参数单卡部署，单张 RTX 5090 超 24 FPS，一分钟流式推理仅约 6 分钱，消费级显卡即可实时交互。架构轻量优雅：动作控制链路不足 10M 参数（sin/cos 编码 + 因果卷积 → 逐帧 action residual），双控制链共同进入 Causal DiT，设计简洁高效。真开源：Apache 2.0 协议，代码 + 权重 + 部署文档全套开放，商用无门槛，基座 Wan2.2-TI2V-5B 成熟可靠。Zing-0.5的项目地址Github仓库：https://github.com/seedleap/zing-world-modelHuggingFace模型库：https://huggingface.co/seedleap/zing-0.5ModelScope：https://modelscope.cn/models/seedleap/Zing-0.5Zing-0.5的同类竞品对比对比项Zing-0.5（Loopit/SeedLeap）HiDream-O1-World（HiDream.ai）LingBot-World v2（蚂蚁集团）WBench 总排名第 2（实时榜第 1）第 3第 5综合平均分81.080.979.4画面质量（Quality）80.681.0（三者最高）81.8（三者最高）场景设定（Setting）77.882.276.8交互能力（Interaction）84.2（三者最高）80.082.8一致性（Consistency）88.5（三者最高）88.086.5物理规律（Physical）73.8（三者最高）73.369.1实时性实时（单卡 5090 超 24 FPS）未标注实时标签标注 “fast”（快速）交互类型action + 文本联合控制（可中途改写世界）动作控制动作控制参数规模5B（基座 Wan2.2-TI2V-5B）未公开未公开开源情况开源（Apache 2.0，代码+权重）未标注开源开源核心特点边操控边文字改写世界，新指令进入当前世界不重置场景画面质量与场景设定最强速度快，交互能力较好推理成本约 6 分钱/分钟流式生成未公开未公开产品落地2 周内上线 Loopit APP 海外版——Zing-0.5的应用场景AI 互动娱乐／AI 游戏：玩家用键盘操控角色在生成世界中探索，随时用语言改写剧情与环境（如下雪、召唤怪物），实现”人人可玩的生成式游戏”。“可以玩的抖音”式内容消费：短视频从”看”变成”玩”，每个用户在同一段内容里玩出不同走向，改写内容平台形态。UGC 互动内容创作：创作者生成可交互视频世界供粉丝体验，观众不再是旁观者而是参与者，衍生全新内容品类。游戏原型快速验证：游戏团队无需搭建引擎场景，用世界模型快速生成可交互 Demo，低成本验证玩法与关卡创意。虚拟直播／虚拟主播：主播实时驱动虚拟世界，观众通过弹幕文字参与改写场景与事件，增强直播互动性。# 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2026 年 AI 生成图文店门头效果图 - 7 款工具深度横评下一篇Miya - 时域科技推出的 AI 音乐应用，画面可随歌词律动相关文章Giga视频超级工厂 – AI短视频批量制作平台，提供智能化解决方案AI小集7免费 AI 外观专利检索神器，跨境卖家必备刚需AI小集3SCAIL – 智谱AI开源的影视级角色动画生成框架AI小集3Jellypod – AI播客制作平台，支持多种内容格式自动生成播客脚本AI小集3ASAM – vivo公司推出的AI图像分割模型AI小集2Maestro – 开源的端到端自动化测试框架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搜索妙呀堆友Agent绘蛙AI美图设计室updream办公小浣熊秒哒最新收录DeepSkillAlphaVowXCellMetaDigChatArtLexcat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秒前WeMM-Embedding – 腾讯开源通用多模态Embedding模型2小时前FireRedAudio – 小红书 FireRed 推出的通用音频语言模型2小时前MiniMax H3