[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-gemini-4-post-training":3,"topics-all":38,"news-related-5de53eff-f921-49ac-9e11-5849b4462295":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},"5de53eff-f921-49ac-9e11-5849b4462295","Gemini 4 进入后训练,谷歌跳过 3.5 Pro","谷歌 DeepMind 负责人 Kavukcuoglu 周三确认,新一代旗舰模型 Gemini 4 已进入后训练早期阶段,首个版本将远早于年底发布,之后持续快速迭代;原定六月的 3.5 Pro 被跳过,由 Flash 系列承担高频更新;新模型已在内部运行编程工具,并与自研芯片路线形成协同,用于加快追赶领先旗舰。","9 月 24 日,谷歌 DeepMind 负责人 Koray Kavukcuoglu 在 The Information 举办的活动上确认:下一代旗舰模型 **Gemini 4** 已进入\"后训练\"早期阶段,希望\"远早于\"今年年底发布首个版本([财联社](https:\u002F\u002Fwww.cls.cn\u002Fdetail\u002F2491924)、[新浪财经](https:\u002F\u002Ffinance.sina.com.cn\u002Fstock\u002Fusstock\u002Fc\u002F2026-09-24\u002Fdoc-inisxhnz8166381.shtml))。\n\n## 后训练阶段意味着什么\n\n后训练是在基础模型完成大规模预训练之后,进一步优化模型,使其在可靠性、行为一致性和实际任务表现上达到可发布水平。据新浪财经报道,谷歌已看到内部测试结果,计划尽快释放一个早期后训练版本,而不是等所有优化结束后一次性发布,之后持续快速迭代。这意味着 Gemini 4 的发布策略更接近“先推可用版本、再持续升级”,以缩短从训练完成到上线的时间。\n\n## 跳过 Gemini 3.5 Pro\n\n谷歌原本计划今年 6 月推出 Gemini 3.5 Pro——CEO Sundar Pichai 曾在 5 月的 I\u002FO 大会上宣布这一时间表,但该模型最终没有出现。Kavukcuoglu 解释,在 Gemini 3 和 3.1 发布之后,谷歌选择“退一步”,把更多精力放在 Flash 系列上,以加快团队学习和迭代。他没有明确 3.5 Pro 是否正式取消,但确认当前重点已转向 Gemini 4。\n\n## 内部已经上岗\n\nGemini 4 不只停留在实验室基准测试里。据新浪财经报道,谷歌正为 Gemini 4 开发安全防护机制并进行安全测试,工程师已在内部使用该模型运行 Antigravity AI 编程工具。把模型提前投入真实工程环境,可以在正式发布前暴露复杂代码任务、工具调用以及长链条任务执行的问题。\n\n## 追赶背景与 TPU 协同\n\n这轮提速的背景是前沿模型竞争加剧:Anthropic 今年春季推出新一代 Mythos 模型系列并在本月继续升级,OpenAI 则在 9 月初开始推出 GPT-6 系列;报道认为两家最新旗舰在复杂推理上已超过谷歌现有最强产品。被问及谷歌是否落后时,Kavukcuoglu 表示对 DeepMind 团队保持信心。\n\n另一个容易被忽略的维度是硬件协同。据新浪财经报道,谷歌今年开始直接向客户销售 TPU,而 DeepMind 会让硬件团队提前了解未来模型的算力需求,据此规划未来两到三代 TPU 的设计。模型定义芯片、芯片反哺模型,这个闭环是纯模型公司不具备的。\n\n## 怎么看\n\n三点判断:\n\n1. **跳代是资源集中的信号。** 放弃 3.5 Pro、直接押注 Gemini 4,说明谷歌把旗舰研发对准代际跃迁,而让 Flash 系列承担高频迭代——旗舰与轻量线的分工比\"憋大招\"更清晰。\n2. **“先可用、再迭代”会压缩选型窗口。** 早期后训练版本一旦对外,下游面对的将不是一次性版本对比,而是持续滚动的模型状态,评估节奏必须跟上。\n3. **AGI 叙事让位于可信智能体。** Kavukcuoglu 淡化\"是否达到 AGI\"的讨论,强调能否构建可以信任的智能体——这和行业重心从刷榜转向稳定执行长任务的方向一致。\n\n所以呢:如果你的技术选型表里留着\"等 Gemini 3.5 Pro\"的位置,现在该改写成\"年底前评估 Gemini 4 早期版本 + 快速迭代\"——谷歌自己都已经把中期版本从路线图上拿掉了。","https:\u002F\u002Fwww.cls.cn\u002Fdetail\u002F2491924","b1c980f4-2409-4767-adb6-a7ccdd79f990",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"a9524a82-a7c5-4daa-bb4b-a7ee77bb0b94","gemini",{"id":19,"name":20,"slug":20,"description":14,"color":14},"8cf7490f-2449-4ba7-be19-61befa0d92b4","google",{"id":22,"name":23,"slug":23,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"ed5be889-2c5e-4eaf-a0be-b72a6dcf85de","en","Gemini 4 Enters Post-Training, Google Skips 3.5 Pro","Google says Gemini 4 is in early post-training, targets release well before year-end; the planned 3.5 Pro was skipped, Flash models iterate fast.","On September 24, Google DeepMind chief Koray Kavukcuoglu confirmed at an event hosted by The Information that the next flagship model, **Gemini 4**, has entered the early \"post-training\" phase of development, and that Google hopes to release the first version \"well before\" the end of the year ([CLS](https:\u002F\u002Fwww.cls.cn\u002Fdetail\u002F2491924), [Sina Finance](https:\u002F\u002Ffinance.sina.com.cn\u002Fstock\u002Fusstock\u002Fc\u002F2026-09-24\u002Fdoc-inisxhnz8166381.shtml)).\n\n## What post-training means\n\nPost-training is the stage where a base model, after large-scale pre-training, is further optimized so that its reliability, behavioral consistency, and real-task performance reach a shippable level. According to Sina Finance, Google has already seen internal test results and plans to ship an early post-training build rather than waiting until every optimization is done, then iterate rapidly afterwards. In other words, the Gemini 4 release strategy is closer to \"ship a usable version first, keep upgrading\", shortening the gap between training completion and launch.\n\n## Skipping Gemini 3.5 Pro\n\nGoogle had originally planned to launch Gemini 3.5 Pro in June — CEO Sundar Pichai announced that timeline at I\u002FO in May, but the model never appeared. Kavukcuoglu explained that after Gemini 3 and 3.1, Google chose to \"take a step back\" and focus more on the Flash family to speed up learning and iteration. He did not say whether 3.5 Pro was officially cancelled, but confirmed the focus has now shifted to Gemini 4.\n\n## Already on internal duty\n\nGemini 4 is not just sitting in lab benchmarks. According to Sina Finance, Google is developing safety guardrails and running safety tests for the model, and engineers already use it to run the internal Antigravity AI coding tool. Putting the model into a real engineering environment early exposes problems in complex coding tasks, tool use, and long-horizon execution before the public release.\n\n## Catch-up context and the TPU loop\n\nThe acceleration comes amid renewed frontier competition: Anthropic launched its new Mythos model family this spring and kept upgrading it this month, while OpenAI began rolling out its GPT-6 series in early September; reports suggest both companies' latest flagships already surpass Google's strongest current models in complex reasoning. Asked whether Google is behind, Kavukcuoglu said he retains full confidence in the DeepMind team.\n\nThe other easily missed dimension is hardware. Sina Finance reports that Google started selling TPUs directly to customers this year, and that DeepMind's model roadmap feeds the hardware team early signals about future compute needs, letting it plan the next two to three TPU generations accordingly. Models define chips, chips feed back into models — a structural loop that pure model labs do not have.\n\n## How to read it\n\nThree takeaways:\n\n1. **Skipping a generation is a resource-concentration signal.** Dropping 3.5 Pro and betting directly on Gemini 4 means the flagship line aims at generational leaps while the Flash family handles high-frequency iteration — a cleaner division of labor than \"saving up for one big release\".\n2. **\"Usable first, iterate fast\" compresses evaluation windows.** Once early post-training builds go public, downstream teams face a continuously rolling model state rather than one-off version comparisons; procurement and benchmarking cadence has to keep up.\n3. **AGI narrative gives way to trustworthy agents.** Kavukcuoglu downplayed the \"has AGI been achieved\" debate and emphasized building agents you can actually trust — consistent with the industry shift from leaderboard chasing to reliable long-horizon task execution.\n\nSo what: if your technology shortlist still has a slot for \"waiting on Gemini 3.5 Pro\", rewrite it as \"evaluate an early Gemini 4 before year-end, then track rapid iterations\" — Google itself has already removed the mid-cycle version from its own roadmap.","gemini-4-post-training","2026-09-24T13:20:00Z","2026-09-24T13:12:52.982447Z","2026-09-24T13:12:52.982455Z",true,"agent",563,[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},"e0a484e4-41e0-4f91-9b9b-a196bbdcf3ba","Gemini 3.8 Flash 双发:同价升级 + Cyber 走可信项目 Fairwind","gemini-3-8-flash-cyber-fairwind-launch","2026-09-03T03:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"f093935e-e23c-4766-88d6-9302a370be2a","Gemini 套餐洗牌:免费版限 Flash-Lite","gemini-deep-think-ai-pro-tier-restructure","2026-10-09T00:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"90606043-6747-4b31-b6a5-76d38c649d1c","Nano Banana 2.1:价格砍半,跑分压过Pro","nano-banana-2-1-halves-image-price","2026-10-08T07:10:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"7a85c2fe-b439-46e7-b531-282e3d787f5a","Googlebook 上架第一天:Magic Pointer 把 Gemini 拉到光标","googlebook-magic-pointer-cursor-launch","2026-10-05T02:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"260aa193-a883-4db9-9e05-7bf3c6f14f2f","Gemini 4 Argon 发布:输出上限 100 万 token,首批给安全防御者","gemini-4-argon-release","2026-10-01T13:11:08+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"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"]