{
 "generated": "2026-11-05",
 "sha": "707df050bf85",
 "node": {
  "id": "ep:104",
  "type": "Episode",
  "name": "EP.104",
  "slug": "104",
  "ep": 104,
  "date": "2026-09-25",
  "page": "/buzzwords/104/",
  "n_models": 19,
  "window_from": "2026-09-18",
  "window_to": "2026-09-25",
  "window": "2026-09-18 → 2026-09-25",
  "meta": "窗口内发布或开放的主要新模型，按类型分组；每个模型列开发者、发布日、规模、可得性和许可、要点与一手链接。带“边界日”或“背景”标记的，是本期的故事线和快讯里讲到、但发布日在窗口之前的模型。",
  "intro": "本期是补录（2026-10-23 核对）：模型清单按本期页面当时写到的内容定，规格、价格、许可和可得性是 10-23 读取的一手页，可能与当期不同（例如权重后来才放出的）。发布日以官方公告或模型卡为准，不是 HF 仓库的创建日。要点后括号里标了信息来自“厂商自述”“模型卡”还是“独立”：厂商自述的基准本刊没有重跑，模型卡指规格和许可类事实来自 HF 卡或官方文档本身，独立指第三方榜单或媒体实测。各类按窗口内的在前、边界日和背景的在后排。",
  "excluded": "没有列入的，以及原因。窗口之后才发布，只出现在本页发布后补写的“10-08 补充”里：Fastino 的 GLiDE（09-30）、Gemini 4 Argon（09-30）、Mistral Large 4（10-06，台账里已有 mistral-large-4）、地平线 CogWAM（GitHub 09-27、权重 10-01）、Databricks ai_decide、MiMo-V2.6 的 MOPD 检查点（09-27）；OpenAI 的 GPT-6 Cyber 只是据 Fortune 报道在预览，官方页没有。更新版而非新模型：Qwen3.8-Max（阿里 09-22 云栖大会称“更新版”的 Artificial Analysis 分数从 40 升到 45，是自报；原版更早发布，本页只引这条自报，没有新的发布稿）。不是新模型：SemIf（原 OpenJev，用冻结的 Qwen3.5-4B 直接读 logits 的方法和浏览器实现，Tunguz 把它与 Jev 对照）；Asymptote Labs 的 Beacon、Jevable 目录、Simon Willison 的 llm-typesafe 插件；Meta Muse 智能体与 Muse Charm 硬件；阿里真武 V900、高通 2nm 芯片、宇树 Dex5-S；Intrinsic Core、Isaac ROS 5.0、InstinctFlash 等软件栈。研究预印本或没给名字：Stanford 的 Minerva 基因组语言模型（bioRxiv，没放权重）、PhAI Labs 为 ScienceIDE 发的“三个配套模型”（页面没给名字，也没在其项目页找到）；Anthropic 的酶系统发现用的是已有的 Claude Mythos 5。窗口之前、只作背景或对比：GLM-5.3 与 GLM-5.3-Flash（EP.101 已列，本期只引开源榜名次与 NIST 评估）、Kimi K3、Nemotron 3 Ultra、Thinking Machines Inkling（只在 Artificial Analysis 榜单里出现）、NVIDIA Cosmos 3（05-31，Nano 16B / Super 64B）、LeWorldModel（03-13，论文）、Claude Fable 5.1、Mythos 5、Opus 4.8、GPT-6 Astra（对比或回退项）。日期口径：Step 5 Preview 本页写 Artificial Analysis 登记 09-18 发布、10-15 开放权重，HF 仓库 09-20 就有上传，EP.108 的 excluded 写“9-20 放权重”，已在可得性里写明三个说法。 本期是补录（2026-10-23）：只收本期页面里提到的模型发布，窗口内页面没提到的不在此列。",
  "n_in_window": 10,
  "cat_notes": [
   {
    "category": "open-llm",
    "text": "CUA-S1-FORMS 是只有 70 万参数的“判定模型”，放在这里是因为分类表里没有对应类；EP.108 已另设“决策模型”一类，建议把它和 Jev 一起挪过去。"
   },
   {
    "category": "closed-llm",
    "text": "Jev 是不生成文字的“判定模型”，暂放此类；建议与 EP.108 的“决策模型”类合并。"
   },
   {
    "category": "video-world",
    "text": "这一组里 FLUX 3 Action 和 Helix 2.5 是机器人动作模型（世界-动作 / 视觉-语言-动作），暂放最接近的一类；建议以后单列“机器人 / 具身”一类。"
   }
  ],
  "title": "评估员开张、Agent闯进真公司、算力开始涨价",
  "oneliner": "Anthropic 签下第一家驻场评估方埃森哲、钱自己付，四天后与 OpenAI 同日发新模型还降价；Gemini 测试中闯进三家真公司、白帽用 Claude 打进 OpenAI、AI 假情报让美军飞机升空；Nebius 三个月第二次涨价、AMD 破万亿，得州加州弗吉尼亚同周卡住数据中心",
  "template": "new",
  "status": "published",
  "n_must_index": 20,
  "kind": "full",
  "kind_zh": "完整页",
  "entry": "agent",
  "by": "episode-import",
  "at": "2026-11-04T16:15:12",
  "n_stories": 13,
  "n_topics": 10,
  "n_items": 44,
  "n_sections": 5,
  "n_recap": 5,
  "n_must": 19,
  "n_dispatch": 11,
  "n_howto": 4,
  "n_outlook": 5,
  "degree": 0,
  "n_entities": 93,
  "scale": "13 个故事线 / 专题 · 44 个条目 · 讲到 93 个实体"
 },
 "page": "https://xiaopingfeng.com/graph/e/ep/104/",
 "out": [],
 "in": [
  {
   "type": "part_of",
   "sentence": "喊慢第二周：评估员拿到了第一单，新模型照发还降了价属于EP.104",
   "direction": "in",
   "id": "story:104-story-a",
   "name": "喊慢第二周：评估员拿到了第一单，新模型照发还降了价",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Agent 走出了沙箱：闯进真公司、打进 OpenAI、差点让军机登船属于EP.104",
   "direction": "in",
   "id": "story:104-story-b",
   "name": "Agent 走出了沙箱：闯进真公司、打进 OpenAI、差点让军机登船",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "卖算力的在涨价，建数据中心的被三个州同时卡住属于EP.104",
   "direction": "in",
   "id": "story:104-story-c",
   "name": "卖算力的在涨价，建数据中心的被三个州同时卡住",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "不会聊天的模型，一周内被人拿去真用了：判定模型成了一个品类属于EP.104",
   "direction": "in",
   "id": "story:104-dg-decision-models",
   "name": "不会聊天的模型，一周内被人拿去真用了：判定模型成了一个品类",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Jev 带来的机会：认真查了一圈，目前还没人接属于EP.104",
   "direction": "in",
   "id": "story:104-dg-jev-opportunity",
   "name": "Jev 带来的机会：认真查了一圈，目前还没人接",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "开源榜前五全是中国模型，第一名一周内换了人属于EP.104",
   "direction": "in",
   "id": "story:104-dg-china-open",
   "name": "开源榜前五全是中国模型，第一名一周内换了人",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "AI 做科学往前走了一步：进了湿实验室，找到了一个新酶，也进了四分之一数学论文的致谢栏属于EP.104",
   "direction": "in",
   "id": "story:104-dg-ai-science",
   "name": "AI 做科学往前走了一步：进了湿实验室，找到了一个新酶，也进了四分之一数学论文的致谢栏",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "唱片公司的新说法：\"洗白\"——旧案只告了 560 首，新案补上 6 万首，再加一个\"v6 也被污染\"的论点属于EP.104",
   "direction": "in",
   "id": "story:104-dg-suno-laundering",
   "name": "唱片公司的新说法：\"洗白\"——旧案只告了 560 首，新案补上 6 万首，再加一个\"v6 也被污染\"的论点",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "机器人软件栈的\"开源时刻\"？同一个生态里三家互相点名的公司，同一天在同一个会上协同发布属于EP.104",
   "direction": "in",
   "id": "story:104-dg-robotics-opensource",
   "name": "机器人软件栈的\"开源时刻\"？同一个生态里三家互相点名的公司，同一天在同一个会上协同发布",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "芯片战的矛与盾：华为把路线图提前的同一周，国会议员喊话收紧\"云服务漏洞\"属于EP.104",
   "direction": "in",
   "id": "story:104-dg-chip-geopolitics",
   "name": "芯片战的矛与盾：华为把路线图提前的同一周，国会议员喊话收紧\"云服务漏洞\"",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "融资链的\"第一个断点\"？Oracle 的不可抗力通知是一场租金起付日之争，监管者和债券买家的话也都没指向它属于EP.104",
   "direction": "in",
   "id": "story:104-dg-financing-chain",
   "name": "融资链的\"第一个断点\"？Oracle 的不可抗力通知是一场租金起付日之争，监管者和债券买家的话也都没指向它",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "世界模型之争：一边说别预测像素，一边说规模还没到头，但没有一把能同时量两边的尺子属于EP.104",
   "direction": "in",
   "id": "story:104-dg-world-model-debate",
   "name": "世界模型之争：一边说别预测像素，一边说规模还没到头，但没有一把能同时量两边的尺子",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "网络安全模型的\"分级放行\"：不是同周发明的，是两家共同的做法——差别在最强那一档留给谁属于EP.104",
   "direction": "in",
   "id": "story:104-dg-cyber-tiered-access",
   "name": "网络安全模型的\"分级放行\"：不是同周发明的，是两家共同的做法——差别在最强那一档留给谁",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "OpenAI 和 Google 都不会公布具名第三方评估机构属于EP.104",
   "direction": "in",
   "id": "item:104-recap-1",
   "name": "OpenAI 和 Google 都不会公布具名第三方评估机构",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "没有中国头部实验室接 Musk 的互测提议；联合声明不写互测属于EP.104",
   "direction": "in",
   "id": "item:104-recap-2",
   "name": "没有中国头部实验室接 Musk 的互测提议；联合声明不写互测",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "OpenAI 会按期回复 Blumenthal，但不公开执行轨迹属于EP.104",
   "direction": "in",
   "id": "item:104-recap-3",
   "name": "OpenAI 会按期回复 Blumenthal，但不公开执行轨迹",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "H.R. 9340 众议院简易程序通过，参议院 09-25 前不跟属于EP.104",
   "direction": "in",
   "id": "item:104-recap-4",
   "name": "H.R. 9340 众议院简易程序通过，参议院 09-25 前不跟",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "09-25 前 EDGAR 上不会出现 Anthropic 的公开 S-1属于EP.104",
   "direction": "in",
   "id": "item:104-recap-5",
   "name": "09-25 前 EDGAR 上不会出现 Anthropic 的公开 S-1",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "NYT 诉 OpenAI 案解封文件：微软内部文档写着\"末日循环\"属于EP.104",
   "direction": "in",
   "id": "item:104-ni-5d1e58",
   "name": "NYT 诉 OpenAI 案解封文件：微软内部文档写着\"末日循环\"",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "网信办征求意见：不得向未成年人提供虚拟伴侣类 AI 服务属于EP.104",
   "direction": "in",
   "id": "item:104-ni-ba228c",
   "name": "网信办征求意见：不得向未成年人提供虚拟伴侣类 AI 服务",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "独立研究者逆向：ChatGPT 广告用一个 cookie 跨站追踪用户属于EP.104",
   "direction": "in",
   "id": "item:104-ni-488d9b",
   "name": "独立研究者逆向：ChatGPT 广告用一个 cookie 跨站追踪用户",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Grok 4.7 发布：换了更大的基座，价格不变属于EP.104",
   "direction": "in",
   "id": "item:104-ni-41d87d",
   "name": "Grok 4.7 发布：换了更大的基座，价格不变",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Claude Code Projects 改版：从\"文件夹\"变成\"对话\"属于EP.104",
   "direction": "in",
   "id": "item:104-ni-17d89f",
   "name": "Claude Code Projects 改版：从\"文件夹\"变成\"对话\"",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Qwen-Image-2.1：阿里这周第三个模型，统一了生图和图像编辑属于EP.104",
   "direction": "in",
   "id": "item:104-ni-aeb2a2",
   "name": "Qwen-Image-2.1：阿里这周第三个模型，统一了生图和图像编辑",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "FLUX 3 Action：图像模型公司跨界做机器人大脑属于EP.104",
   "direction": "in",
   "id": "item:104-ni-2f5f5a",
   "name": "FLUX 3 Action：图像模型公司跨界做机器人大脑",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Qwen3.8-LiveTranslate：同传延迟降到 2.3 秒，这次没放权重属于EP.104",
   "direction": "in",
   "id": "item:104-ni-b3c201",
   "name": "Qwen3.8-LiveTranslate：同传延迟降到 2.3 秒，这次没放权重",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Figure Helix 2.5：没见过的 30 户人家里零样本做家务，成功率 56%属于EP.104",
   "direction": "in",
   "id": "item:104-ni-09536e",
   "name": "Figure Helix 2.5：没见过的 30 户人家里零样本做家务，成功率 56%",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "梁文锋署名新论文：DeepSeek 自家的 Agent 训练沙箱基础设施属于EP.104",
   "direction": "in",
   "id": "item:104-ni-830e08",
   "name": "梁文锋署名新论文：DeepSeek 自家的 Agent 训练沙箱基础设施",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "宇树发布 Dex5-S 灵巧手：22 自由度，和真手同尺寸属于EP.104",
   "direction": "in",
   "id": "item:104-ni-9531fa",
   "name": "宇树发布 Dex5-S 灵巧手：22 自由度，和真手同尺寸",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "用 AIPerf 做大规模 LLM 推理压测↗属于EP.104",
   "direction": "in",
   "id": "item:104-ht-1dadf6",
   "name": "用 AIPerf 做大规模 LLM 推理压测↗",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Chief of Staff 模式：用一个 Claude Code 会话管住一群会话↗属于EP.104",
   "direction": "in",
   "id": "item:104-ht-d5e323",
   "name": "Chief of Staff 模式：用一个 Claude Code 会话管住一群会话↗",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "让 Agent 拿基准测试当反馈，把 Rust 代码调到比主流库还快↗属于EP.104",
   "direction": "in",
   "id": "item:104-ht-63d15a",
   "name": "让 Agent 拿基准测试当反馈，把 Rust 代码调到比主流库还快↗",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Transformers 直接加载 llama.cpp 的 GGUF 量化模型↗属于EP.104",
   "direction": "in",
   "id": "item:104-ht-47eb27",
   "name": "Transformers 直接加载 llama.cpp 的 GGUF 量化模型↗",
   "status": "derived",
   "evidence": []
  },
  {
   "type": "part_of",
   "sentence": "Anthropic 的公开 S-1 会不会在 10 月 2 号前挂出来属于EP.104",
   "direction": "in",
   "id": "item:104-nw-39",
   "name": "Anthropic 的公开 S-1 会不会在 10 月 2 号前挂出来",
   "status": "derived",
   "evidence": []
  }
 ]
}