[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-deepmind-genie-4-native-audio-world-model":3,"topics-all":38,"news-related-1068c0fb-47cd-4a67-a55c-18dc27838561":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"1068c0fb-47cd-4a67-a55c-18dc27838561","Genie 4 发布:原生音频实时模拟,单次可达两分钟","谷歌 DeepMind 发布世界模型 Genie 4 研究预览:该系列首个原生生成音频的实时世界模型,环境音、音乐与可交互对话和画面同时生成。Gemini 3.8 Flash 在回路里做引导代理,Gemini Omni 锚定真实外观。单次模拟最长约两分钟,约为上一代的两倍,一致性仍是待解问题。","10 月 27 日,谷歌 DeepMind 在官方博客发布 Genie 4 的研究预览,署名作者为 Raia Hadsell、Diego Rivas 和 Hang Qi。这是 Genie 系列的新一代模型,前代 Genie 3 发布于 2025 年 8 月。目前只向受信任的测试者开放,之后将进入 Project Genie 的 Google AI Ultra 订阅者,以及 Gemini API 和 Google AI Studio 的早期合作伙伴。\n\n真正的变化不在画面,而在声音。DeepMind 在原文里写明,Genie 4 是他们首个原生生成音频的实时世界模型([官方博客](https:\u002F\u002Fdeepmind.google\u002Fblog\u002Fgenie-4-a-new-generation-of-world-simulation\u002F)):环境音、背景音乐和可交互对话与画面同时生成。财联社的报道用了同样的定性,称它是该系列首款能同步生成声音和画面的产品([财联社报道](http:\u002F\u002Ffinance.eastmoney.com\u002Fa\u002F202610283910639274.html))。网易科技还提到,这些对白并非提前配好、反复播放的固定台词,用户中途改变情节,人物也能顺着新情况接话;物理层面则覆盖空气动力学、碰撞等场景,人物动作可以被实时操控。\n\n## 推理在 Gemini,渲染在 Genie\n\nGenie 4 并不是一个单模型。它把两个家族拼在一起:Gemini Omni 负责把模拟锚定在真实世界的外观与行为上,其他 Gemini 模型负责维持场景的连贯与连续性;真正在回路里当引导代理的是 Gemini 3.8 Flash——它持续观察传入的画面,理解用户目标,把目标拆成步骤序列,再回头核查每一步是否完成。财联社提到,团队其实试过 Argon,但实际驱动实时体验的是较轻量的 Gemini 3.8 Flash。\n\nDeepMind 把这条路线称为混合世界模型:代码与语言模型负责逻辑和结构化推理,像素级模拟负责感官的丰富表征。另有一组实验用 Gemini 4 Argon、Genie 4 和 Gemini Robotics 组合,让 Apptronik 的 Apollo 人形机器人在虚拟棋盘上下棋——那是模拟,不是实体机器人真的学会了下棋。\n\n## 两分钟的天花板\n\n最该记住的数字是时长。财联社报道,模拟环境最长可运行两分钟,约为上一代 Genie 3 的两倍;网易科技援引 TechCrunch 的说法同样是单次模拟约两分钟,该报道并指出今年 1 月上线的 Project Genie 原型每次生成最多 60 秒([网易科技报道](http:\u002F\u002Fstock.10jqka.com.cn\u002Fusstock\u002F20261028\u002Fc681165639.shtml))。\n\n但两分钟不等于两分钟的一致。据财联社,DeepMind 研究工程师 Hang Qi 表示,即便模型能够运行至两分钟的上限,也不意味着场景在整个过程中都能保持完全一致;研究总监 Tejas Kulkarni 则说,要让虚拟世界完全保持一致,即使仅仅维持两分钟,也需要模型真正理解日常世界的物理规律。产品经理 Diego Rivas 的表述更直接:两分钟能做到的事情终究有限,团队仍处于研究探索阶段,还没有真正进入最终产品的开发阶段。官方列出的待解清单还包括视觉细节与物体恒存会随时间漂移、实时音频准确度、空间准确性。生成的音视频全部带 SynthID 隐形水印。\n\n## 预演、原型与训练环境\n\n受邀测试者的用法集中在四处:电影创作者把它当拍摄预演,同一场戏换机位、改台词即时试;摄影记者把一张 2D 档案照片变成可探索场景,用来推演山洪或山火,但补出的画外细节不能被当成历史事实;游戏开发者先用自然语言探索风格与玩法原型,再决定写不写引擎代码;企业伙伴拿它做从零到一的场景生成,谷歌自己的机器人研究也拿它当训练环境。据网易科技,皮查伊在发布帖中把 Genie 4 称为谷歌迄今最逼真、互动性最强的世界模型;同一篇还提到今年 1 月 Project Genie 原型上线后,Unity、Roblox、Take-Two 等游戏股一度下跌。\n\n我的判断:这一步的分水岭不是画面更真,而是官方第一次把推理和渲染摆成两个可替换的部件,世界模型由此更像一份接口约定而不是一件成品。也正因如此,两分钟这个天花板比任何演示都更值得记住:它现在仍是研究预览,离能替代引擎的完整产品还很远。等混合架构真的成为标准,你要买的是推理那一层还是渲染那一层,才是下一个真问题。","https:\u002F\u002Fdeepmind.google\u002Fblog\u002Fgenie-4-a-new-generation-of-world-simulation\u002F","35ce748f-48b7-4638-88ef-effa57a7e749",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"a9524a82-a7c5-4daa-bb4b-a7ee77bb0b94","gemini",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":19,"name":20,"slug":20,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":22,"name":23,"slug":23,"description":14,"color":14},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"d154181f-296c-48c4-98d8-0c931c2b622b","en","Genie 4: native audio, Gemini reasoning, two-minute ceiling","Genie 4 is DeepMind's first real-time world model with native audio: Gemini 3.8 Flash steers the scene, and each run caps near two minutes.","On October 27, Google DeepMind published a research preview of Genie 4 on its official blog, authored by Raia Hadsell, Diego Rivas and Hang Qi. It is the newest model in the Genie line; the previous generation, Genie 3, shipped in August 2025. Access starts with a small group of trusted testers, then rolls out to Google AI Ultra subscribers inside Project Genie and to early access partners of the Gemini API and Google AI Studio.\n\nThe real change is not the picture, it is the sound. DeepMind states plainly in the original post that Genie 4 is its first real-time world model to generate audio natively: ambient noise, music and interactive dialogue are produced alongside the world itself (official blog: https:\u002F\u002Fdeepmind.google\u002Fblog\u002Fgenie-4-a-new-generation-of-world-simulation\u002F). CaiLian Press reports the same characterization, calling it the first release in the series that can generate sound and picture in sync (https:\u002F\u002Ffinance.eastmoney.com\u002Fa\u002F202610283910639274.html). NetEase Technology adds that the dialogue is not a pre-recorded loop — change the plot mid-session and characters answer according to what just happened. On physics, the model covers aerodynamics and collisions, and characters can be steered in real time.\n\n## Reasoning in Gemini, rendering in Genie\n\nGenie 4 is not one model. It stitches together two families. Gemini Omni anchors the simulation in real-world appearance and behavior, while other Gemini models maintain coherence and continuity. The guiding agent actually running in the loop is Gemini 3.8 Flash: it watches incoming frames, works out what the user wants, breaks the goal into a sequence of steps, and reviews the frames to verify that each step succeeded. CaiLian Press notes the team did try Argon, but the real-time experience is actually driven by the lighter Gemini 3.8 Flash.\n\nDeepMind calls this direction hybrid world models: code and language models carry the logical, structured reasoning, while pixel-level simulation carries the rich sensory representation. A separate experiment combines Gemini 4 Argon, Genie 4 and Gemini Robotics to have Apptronik Apollo humanoid play checkers in a virtual board. That is a simulation, not evidence that a physical robot has learned the game.\n\n## The two-minute ceiling\n\nThe number worth remembering is duration. CaiLian Press reports a simulation environment can run up to two minutes, roughly double the previous Genie 3, and NetEase Technology, citing TechCrunch, gives the same single-run figure of about two minutes, against the Project Genie prototype launched in January which generated at most 60 seconds per run (https:\u002F\u002Fstock.10jqka.com.cn\u002Fusstock\u002F20261028\u002Fc681165639.shtml).\n\nBut two minutes does not mean two minutes of consistency. According to CaiLian Press, DeepMind research engineer Hang Qi said that even running to the two-minute upper limit does not mean the scene stays fully consistent throughout, and research director Tejas Kulkarni said that keeping a virtual world fully consistent, even for two minutes, requires the model to genuinely understand the physics of everyday life. Product manager Diego Rivas was blunter: what two minutes can do is ultimately limited, the team is still in research exploration and has not really entered final product development. The official list of open problems also covers visual detail and object permanence drifting over time, real-time audio accuracy, and spatial accuracy. All generated video and audio carry the SynthID invisible watermark.\n\n## Previsualization, prototypes, training environments\n\nTrusted testers are using it in four main ways. Filmmakers treat it as previsualization, re-blocking a scene or rewriting a line and trying it instantly. Photojournalists turn a single 2D archival photograph into an explorable scene to walk through flash floods or wildfires, though the details filled in beyond the photo frame cannot be treated as historical fact. Game developers explore visual style and gameplay prototypes through natural language prompts before deciding what is worth writing into an engine. Enterprise partners use it for rapid zero-to-one scenario generation, and Google's own robotics work treats it as a training environment. According to NetEase Technology, Sundar Pichai described Genie 4 in his launch post as Google's most photorealistic and most interactive world model yet, and the same report notes that after the Project Genie prototype launched in January, game stocks including Unity, Roblox and Take-Two fell.\n\nMy read: the dividing line here is not more photorealism. It is that Google for the first time put reasoning and rendering on the table as two separable, swappable parts, which makes a world model closer to an interface contract than to a finished product. That is exactly why the two-minute ceiling matters more than any demo: this is still a research preview, far from a complete product that could displace an engine. Once the hybrid architecture really does become the standard, the real question will be whether you are buying the reasoning layer or the rendering layer.","deepmind-genie-4-native-audio-world-model","2026-10-27T16:13:00Z","2026-10-28T03:18:36.730370Z","2026-10-28T03:18:36.730378Z",true,"agent",538,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"03148cbf-c932-48d8-b99b-9132a30fb6cd","Nano Banana 2 Lite 与 Gemini Omni Flash 同日上线:Google 多模态 API 进入分层工业化阶段","nano-banana-2-lite","2026-07-01T02:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"4d1b266e-9e0b-4dd7-85b3-adb5a3c94c7e","Gemini Omni 1.2 Flash:0.07美元每秒","gemini-omni-1-2-flash-video-pricing-drafts","2026-11-05T10:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"bef7a773-36fb-406b-949d-8d728b97b242","Matrix-Game 4.0 开源:单卡 60FPS 的实时可玩游戏世界","matrix-game-4-0-open-source-60fps","2026-10-21T13:09:32+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"240d8a4c-c025-4080-be14-06099f650a71","xAI 把视频价格打到底:1080p 蒸馏版 Lite 比 Veo 3.1 便宜 65%","grok-imagine-1-5-lite-cheaper-than-veo","2026-10-10T08:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"d219ead5-ea5d-4aaa-835e-5ec3cf980856","Gemini Live Avatar:模型直出唇同步数字人","gemini-3-8-live-avatar","2026-09-26T17:06:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"bdb6361a-d1da-45bb-8744-cb6922077e14","谷歌Gemini 3.8双TTS:30秒克隆,逐行导演","gemini-3-8-flash-tts-voice-design","2026-09-23T19:10:29+00:00"]