[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-gemini-4-argon-release":3,"topics-all":38,"news-related-260aa193-a883-4db9-9e05-7bf3c6f14f2f":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},"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","谷歌 9 月 30 日发布 Gemini 4 代首个模型 Argon:单次输出上限从 64K 拉到 100 万 token,DeepSWE v1.1 跑分 77.9% 超过 Opus 5.5 与 GPT-6 Astra;首批用户不是开发者,而是 Fairwind 计划的网络安全防御者,官方对其关闭网络防护栏。","谷歌在 9 月 30 日夜里放出了 Gemini 4 代的第一个模型 Argon。三件事值得记住:单次输出上限从上一代的 64K 直接拉到 100 万 token;DeepSWE v1.1 跑出 77.9%;以及最反直觉的一点——首批用户不是开发者也不是订阅用户,而是 Fairwind 计划里的网络安全防御者。\n\n## 赢在长任务,短板也在明面上\n\n按官方公告,Argon 定位是「复杂、长周期工作流」:真实软件工程、法律金融知识工作、网络防御。100 万 token 是输出上限而非上下文窗口,谷歌没有公布上下文长度。跑分上,官方称 DeepSWE v1.1 的 77.9% 是新纪录,自家对比表里 Claude Opus 5.5 为 74.2%、GPT-6 Astra 为 74.1%;Vals Index、Zapier 的 AutomationBench(51.3%,第一)、长视频理解 LVBench(91.7%)同样领先,CWE-bench v1 以 68% 与 Grok 4.7、GPT-6 Astra 并列第一。\n\n但第三方整理的对比表同样显示它输的地方:FrontierSWE v2 的 55.0% 落后 Astra 的 65.5%,Terminal-Bench 4.0 的 57.4% 落后 Opus 5.5 的 66.4%。有媒体直接提醒:这批头条数字全部来自谷歌自家的评测提交。长任务赢、短平快输,与「100 万输出」的产品定位是自洽的。\n\n## 内部战果:量子、内存、C++ 转 Rust\n\n官方博客晒了三个内部案例:帮量子计算研究员优化子程序的时空资源,几分钟内超过已发表基线 40%;一组 Argon 智能体分析全舰队 profiling 遥测后自主施加内存优化,释放超 300 TiB 内存,预计总节省 500 TiB 到 1 PiB;以及把 C\u002FC++ 代码迁往 Rust——libgav1 里 32K 行 SIMD 代码被替换成可自动向量化的安全 Rust,比原 Rust 移植版快 2.7 倍,Fuchsia 的 Zircon 内核 80 万行以上迁移在进行中。注意,这些全是谷歌自述,尚无外部复现。\n\n## 安全牌才是这次发布的叙事主轴\n\nArgon 被训练成能自主发现、验证并修补漏洞,Wiz 已在 Scan for Good 计划里用它免费扫描关键公共基础设施,早期演示中它在全球医院使用的医疗软件里找出一个此前前沿模型都漏掉的高危漏洞。对可信防御者和谷歌内部团队,官方明确将「关闭网络防护栏」以放行全部防御能力。防护侧则包括内部激活监测、间接提示注入鲁棒性(Gray Swan IPI 榜领先)、对思维链与动作的失配监控。发布节奏上,谷歌正参与美国政府的预发布自愿评估流程,后续才轮到付费 API 与 AI Ultra 订阅用户;介绍期价格 2\u002F10 美元每百万 token,到期恢复 4\u002F20,恰好从 Opus 5.5 的一半涨到持平。\n\n一个值得盯的细节:OpenAI 在 DevDay 刚发完 GPT-6.1 Sol,第二天谷歌就用「先给白帽、不打护栏」抢走了安全叙事。当输出上限冲到 100 万 token,真正的问题不再是模型能不能想,而是你敢不敢让它无人值守地一口气想完。","https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Fmodels-and-research\u002Fgemini-models\u002Fgemini-4-argon\u002F","87334439-1f94-401b-aa88-391be6aaaf93",[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},"8cf7490f-2449-4ba7-be19-61befa0d92b4","google",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"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},"bd93d5e2-be2f-4bf4-8bf2-7448abd3b8f4","en","Gemini 4 Argon arrives: 1M-token output, cyber defenders get it first","Gemini 4 Argon ships with 1M-token output and 77.9% DeepSWE v1.1, beating Opus 5.5. First users are Fairwind cyber defenders, with guardrails off.","Google launched Argon, the first model of the Gemini 4 generation, late on September 30. Three things stand out: the single-response output limit jumps from the previous 64K to 1M tokens; it scores 77.9% on DeepSWE v1.1; and most counterintuitively, the first users are not developers or subscribers but trusted cyber defenders in the Fairwind Program.\n\n## Strong on long-horizon work, weaknesses in plain sight\n\nPer the official announcement, Argon targets complex, long-horizon workflows: real-world software engineering, legal and financial knowledge work, and cyber defense. The 1M figure is an output ceiling, not a context window — Google publishes no context length for Argon at all. On benchmarks, Google says the 77.9% on DeepSWE v1.1 is a new state of the art; its own comparison chart puts Claude Opus 5.5 at 74.2% and GPT-6 Astra at 74.1%. It also leads the Vals Index (68.9%), Zapier's AutomationBench (51.3%, rank #1), and long-video understanding LVBench (91.7%), and ties for first on CWE-bench v1 at 68% alongside Grok 4.7 and GPT-6 Astra.\n\nBut a third-party comparison table shows where it loses: 55.0% on FrontierSWE v2 versus Astra's 65.5%, and 57.4% on Terminal-Bench 4.0 versus Opus 5.5's 66.4%. One outlet flatly noted that all four headline numbers come from Google's own benchmark submissions. Winning long-horizon tasks while losing short ones is consistent with the 1M-output positioning.\n\n## Internal wins: quantum, memory, C++ to Rust\n\nThe official blog showcases three internal cases: helping quantum researchers optimize qubit-times-gates resources of subroutines, beating a published baseline by 40% in minutes; a team of Argon agents analyzing fleet-wide profiling telemetry to autonomously apply memory optimizations, freeing over 300 TiB with an estimated 500 TiB to 1 PiB total savings; and migrating C\u002FC++ to Rust — 32K lines of SIMD code in libgav1 were replaced with auto-vectorizable safe Rust, running 2.7x faster than the prior Rust port, with an 800K+ line Fuchsia Zircon kernel migration underway. Note these are all Google's own accounts with no external replication yet.\n\n## The safety narrative is the real story of this launch\n\nArgon was trained to autonomously find, validate, and patch vulnerabilities. Wiz is already using it in its Scan for Good initiative to scan critical public infrastructure for free; in an early demonstration it uncovered a critical vulnerability in healthcare software used by hospitals worldwide that previous frontier models had missed. For trusted defenders and internal teams, Google explicitly releases Argon without cyber guardrails to unlock full defensive capability. On the protection side: internal-activation monitoring, leading prompt-injection robustness on Gray Swan's IPI benchmark, and misalignment monitoring over chain-of-thought and actions. Google is engaged in the U.S. government's voluntary pre-release evaluation process; paid API and AI Ultra subscribers come later. Introductory pricing is $2\u002F$10 per million tokens, reverting to $4\u002F$20 — from half of Opus 5.5's price to parity.\n\nOne detail worth watching: OpenAI shipped GPT-6.1 Sol at DevDay, and a day later Google seized the safety narrative with \"white hats first, guardrails off.\" When the output ceiling hits 1M tokens, the real question is no longer whether the model can think, but whether you dare let it think unsupervised in one go.","gemini-4-argon-release","2026-10-01T13:11:08Z","2026-10-01T13:11:12.148292Z","2026-10-01T13:11:12.148300Z",true,"agent",303,[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},"039ff515-68e7-4f11-866a-1da97e26eb45","Gemini 3.8 Live 拿下 S2S 实时语音榜第一","gemini-3-8-live-voice-s2s-number-one","2026-09-15T17:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"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":70,"title":71,"news_slug":72,"published_at":73},"16856034-439d-4915-aed4-80b42ae09c68","Gemini 3.7 Flash：FrontierCode 43.6%，价格腰斩","gemini-3-7-flash-coding-agent-fast","2026-08-13T09:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"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":80,"title":81,"news_slug":82,"published_at":83},"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":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"]