[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-ox-alpha-glm-5-3-stealth-zhipu":3,"news-related-2fc64783-8b2a-49a3-939b-edf02bff3622":38},{"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},"2fc64783-8b2a-49a3-939b-edf02bff3622","Ox Alpha 指纹指向 GLM-5.3:OpenRouter 的 1M 上下文隐身模型可能是智谱","8 月 20 日 OpenRouter 与 OpenCode 上线匿名模型 Ox Alpha,1M 上下文、文本+图像+视频输入、独立指纹测试在 25 个提示词与 4 段视频上与智谱 GLM-5.3 \u002F GLM-5V-Turbo 高度吻合,智谱未公开确认。","8 月 20 日,OpenRouter 与 OpenCode 同时上线一个代号 Ox Alpha 的匿名大模型预览。它以 1,048,576 token 的上下文窗口、文本\u002F图像\u002F视频输入、最大 131,072 token 输出,把\"推理模型\"和\"长程 Agent 工作流\"的接口规格直接拉到了当前第一梯队的位置。OpenCode 同期声明该路线对全部用户免费一周、并把每日算力容量定为 100 万亿 token;OpenRouter 上的 prompt 与 completion 价格同样标为零。问题在于:没人署名。OpenRouter 把它挂在通用名 \"Stealth\" 提供商之下,OpenCode 直接称之为\"the stealth model\",Zhipu AI、OpenRouter、OpenCode 三方至今没有给出任何一方的公开确认。\n\n## 独立指纹给出了一个指向 Zhipu 的答案\n\n真正让这件事进入技术圈视野的,是 OpenRouter 用户社区在 36 小时内完成的两轮指纹测试。第一轮由 @aitrackerbot 主导,在 25 条多样化提示词上对比候选模型的原生 token 数。结论是:Ox Alpha 的 token 数与 GLM-5.3 完全一致,仅存在一个恒定的 +75 token 偏移——后者被研究者归因于 OpenRouter 或上游包装层加入的系统提示,与底座身份无关。第二轮把维度升到视频编码器。Ox Alpha 在 4 段受控视频上对帧采样、~147 token\u002F秒 的时长缩放、按帧分辨率的扩展三项独立设计选择,与 Zhipu 自家 GLM-5V-Turbo 的实现逐 token 重合。三个独立设计同时撞车的概率,远低于单个 token 数对得上。同源社区 huggingnews 也在同一时间独立得出\"指纹指向 GLM 5.3 视觉变体\"的判断。\n\n研究者把置信度设在 ~90%,措辞为\"强烈指向但官方未确认\"。Zhipu 至今保持沉默,既没有承认也没有否认;OpenRouter 与 OpenCode 同样拒绝在记录中点名任何实验室。\n\n## 排除项与\"为什么是隐身\"的逻辑\n\n同一轮测试也排除了两个流传最广的替代猜测。小米 MiMo v2.5 的音频端点行为与 Ox Alpha 不同——后者在所有路由上都拒绝音频输入,行为与 GLM-5V 系列对齐,而 MiMo v2.5 原生支持音频。Qwen 3.8 Max 与 GLM-4.6V 也跑过同一套指纹,tokenizer 与视频编码器签名都与 Ox Alpha 不匹配。aiwiki 的复核则把另一组关键事实钉死:OpenRouter 模型 API 把 Ox Alpha 的 tokenizer 标为 \"Other\"、知识截止留空、量化方式 unknown、Hugging Face 仓库 ID 缺失;任何一项单独看都不能证明身份,但合在一起说明这不是一次普通的命名疏忽,而是一次刻意的隐身。\n\n这恰好契合今年早些时候 Hunter Alpha、Healer Alpha 两次小米 MiMo 隐身预览的剧本:在 OpenRouter 这种中立平台上以匿名形式放出、收集生产级流量与失败案例,然后在窗口期结束后再以官方身份正式登场。区别是这一次的指纹指向的不是 MiMo,而是 GLM。\n\n## 不该被忽略的几条边界\n\n值得注意的是,这条\"Ox Alpha = GLM-5.3\"的结论有几个公开的边界。第一,DeepSWE 上有用户用 10 道题做过非受控跑分,宣传位次\"领先 Fable 与 GPT-5.6 Sol\"——但 aiwiki 已指出样本量太小、缺任务 ID 与种子、缺完整轨迹,不能视作官方榜单成绩。第二,OpenRouter 与 OpenCode 各自的隐私条款不一致:OpenRouter 路径下 EULA 授权数据用于训练与改进,OpenCode Zen 才是真正的\"零保留\"——这意味着开发者用 Ox Alpha 跑真实业务时不能把\"两条路径都匿名\"当作统一承诺。第三,1M token 是容量而非有效推理深度:\"Lost in the Middle\" 与 RULER 都未对 Ox Alpha 做受控长上下文测试,长文档中段的检索与多跳推理能力仍是黑盒。\n\n## 所以呢\n\n如果最终公开身份确认是 Zhipu,这意味着头部中国实验室愿意用\"先在海外中立平台裸奔一周、收集真实流量、再以官方身份复活\"的方式做压力测试,与一年前 DeepSeek V3 走过的路径有显著不同——后者更倾向于发论文,而前者直接用生产请求做\"无标签基准\"。无论指纹结果如何,这件事本身的\"隐身→证据链→等待官方确认\"三段式流程,本身就是 2026 年开源大模型生态的一堂新公开课:模型署名不再是发布的前提,但署名迟早必须到来。","https:\u002F\u002Fwww.explainx.ai\u002Fblog\u002Fox-alpha-what-we-know-mystery-ai-model-august-2026","b093fba2-b9ec-492e-acd6-ff6018ef4cfb",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"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},"b1853a5a-d940-42b7-94f9-0488ee3f2cf7","new-model",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"0b90ac3a-5aa8-47bd-bdc9-ba842e3c8b23","en","Ox Alpha fingerprints point to GLM-5.3: Zhipu likely behind OpenRouter's 1M-context stealth model","On Aug 20, OpenRouter and OpenCode launched an anonymous model called Ox Alpha with a 1M context, text\u002Fimage\u002Fvideo input, and a tokenizer + video-encoder fingerprint match to Zhipu's GLM-5.3 and GLM-5V-Turbo across 25 prompts and 4 videos; Zhipu has not publicly confirmed ownership.","On August 20, 2026, OpenRouter and OpenCode simultaneously released a stealth preview of a large model codenamed Ox Alpha. With a 1,048,576-token context window, text\u002Fimage\u002Fvideo input, and a maximum 131,072-token output, it pulled the interface specs for a \"reasoning model\" and \"long-horizon agent workflow\" up to the current top tier. OpenCode announced the route is free to all users for one week and pegged daily compute capacity at 100 trillion tokens; on OpenRouter, the prompt and completion prices are also listed at zero. The catch: no one signed it. OpenRouter listed it under the generic \"Stealth\" provider, OpenCode simply called it \"the stealth model\", and as of writing, Zhipu AI, OpenRouter, and OpenCode have issued no public confirmation from any of the three parties.\n\n## An independent fingerprint points at Zhipu\n\nWhat actually pushed the story into the technical community's view were two fingerprinting passes that the OpenRouter user community completed within 36 hours. The first pass, led by @aitrackerbot, compared native token counts across 25 diverse prompts against the candidate models. The result: Ox Alpha's token counts match GLM-5.3 exactly, with only a constant +75-token offset—which the researchers attribute to a system prompt added by OpenRouter or an upstream wrapping layer, not to the underlying identity. The second pass raised the dimension to the video encoder. On four controlled videos, Ox Alpha matched Zhipu's own GLM-5V-Turbo implementation token-for-token across three independent design choices: FPS-invariant frame sampling, ~147 tokens\u002Fsecond duration scaling, and per-frame resolution scaling. Three independent design choices colliding at once is far less likely than a single token count lining up. The HuggingNews community independently reached the same conclusion around the same time, pointing at a \"GLM 5.3 vision variant\".\n\nResearchers set confidence at ~90%, with the wording \"strongly points to but officially unconfirmed.\" Zhipu has stayed silent, neither confirming nor denying; OpenRouter and OpenCode likewise refuse to name any laboratory on the record.\n\n## What was ruled out—and the logic of going stealth\n\nThe same test pass ruled out the two most-circulated alternative theories. Xiaomi's MiMo v2.5 differs from Ox Alpha in audio-endpoint behavior—Ox Alpha refuses audio input on every route, behaving like the GLM-5V family, whereas MiMo v2.5 natively accepts audio. Qwen 3.8 Max and GLM-4.6V were also put through the same fingerprint and produced tokenizer and video-encoder signatures different from Ox Alpha. A separate AI Wiki cross-check pins down another set of facts: OpenRouter's model API lists Ox Alpha's tokenizer as \"Other\", leaves the knowledge cutoff empty, marks quantization as unknown, and provides no Hugging Face repo ID. None of those alone proves identity, but taken together they suggest this is not ordinary naming carelessness, but a deliberate stealth release.\n\nThat fits the playbook from earlier this year with Hunter Alpha and Healer Alpha, two Xiaomi MiMo stealth previews: drop an unnamed model on a neutral platform like OpenRouter, collect production traffic and failure cases, then reappear under an official identity once the preview window closes. The difference here is that the fingerprint points not at MiMo but at GLM.\n\n## Borders the conclusion should not cross\n\nThe \"Ox Alpha = GLM-5.3\" conclusion has a few public borders worth flagging. First, an informal DeepSWE run of ten tasks circulated with a headline placement \"ahead of Fable and GPT-5.6 Sol\"—but AI Wiki has noted the sample size is too small, lacks task IDs and seeds, and lacks full trajectories, so it cannot be treated as an official leaderboard score. Second, OpenRouter and OpenCode have inconsistent privacy terms: under the OpenRouter route, the EULA authorizes data for training and improvement; only OpenCode Zen is genuinely \"zero retention\"—so developers running real workloads on Ox Alpha cannot treat \"both routes are anonymous\" as one unified promise. Third, 1M tokens is capacity, not effective reasoning depth: neither \"Lost in the Middle\" nor RULER has run controlled long-context tests on Ox Alpha, and retrieval plus multi-hop reasoning over the middle of long documents remains a black box.\n\n## So what\n\nIf the public identity ends up confirmed as Zhipu, this means a top-tier Chinese lab is willing to use a \"bare-run-for-a-week on an overseas neutral platform, collect real traffic, then resurface under an official identity\" pattern as a stress test—markedly different from the path DeepSeek V3 walked a year ago, which leaned more toward papers. This time the lab is using production requests directly as an unlabeled benchmark. Whatever the fingerprint result, the three-act flow—stealth → evidence chain → waiting for official confirmation—is itself a new open lecture in the 2026 open-weight model ecosystem: model attribution is no longer a precondition for release, but attribution will eventually have to arrive.","ox-alpha-glm-5-3-stealth-zhipu","2026-08-22T14:00:00Z","2026-08-22T09:04:21.722168Z","2026-08-22T09:04:21.722176Z",true,"agent",212,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"c049b416-8dac-4d74-b001-ec32c6b49cf3","商汤日日新 SenseNova-U1 Pro 曝光：把「理解·生成·行动」原生统一塞进一个基座，7 月邀测","sensetime-sensenova-u1-pro-unified-multimodal","2026-06-25T18:00:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"00c0b8cc-729a-4daa-b529-ee4313621458","阿里 Qwen-Robot Suite 三连发：把「导航-操作-世界模型」打通成一套具身栈","qwen-robot-suite-nav-manip-world","2026-06-17T02:00:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"a0f1a530-0c0a-423d-8721-69fe339ad118","星火 X2-VL 押注「具身大脑」:从 293B MoE 到多模态感知的国产化闭环","iflytek-x2-vl-293b-moe-embodied-domestic","2026-06-13T12:00:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"caa54bff-d57a-411c-9dae-43f1d4d46875","DeepSeek V4 重磅登场：长期记忆技术突破重塑AI能力边界","deepseek-v4-engram-ltm-long-term-memory-87pct","2026-04-22T07:05:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"8bd5a96a-b85b-4db6-ad54-a2c311867178","字节跳动被曝训练10万亿参数超大模型：对标Anthropic Mythos,中国LLM进入\"10T俱乐部\"前夜","bytedance-10-trillion-parameter-model","2026-08-07T09:11:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"a151db0c-d832-4df2-ac03-2d4e58b26e99","Kimi K3 跑通 MiniTriton:Moonshot 让 LLM 第一次从零编译出自己的 GPU 编译器","kimi-k3-minitriton-gpu-compiler","2026-07-26T14:00:00+00:00"]