{"slug":"longcat-2-5---1-6-7db7f191","url":"https://silinex.xyz/wire/longcat-2-5---1-6-7db7f191","json_url":"https://silinex.xyz/wire/longcat-2-5---1-6-7db7f191.json","md_url":"https://silinex.xyz/wire/longcat-2-5---1-6-7db7f191.md","headline":"Longcat 2 5 1 6 — brief from IT Home","headline_zh":"美团龙猫开源 LongCat-2.5 模型：总参数 1.6 万亿，支持百万上下文","title":"Longcat 2 5 1 6 — brief from IT Home","summary":"Paraphrased Silicon News Wire brief filed from IT Home. Open the Chinese edition (?lang=zh) for the full editorial rewrite.","summary_zh":"综合 IT之家 等公开报道整理：11 月 2 日，美团龙猫发布面向长程智能体任务的 原生多模态模型 LongCat-2.5 。","date_published":"2026-11-09T14:16:38.000+00:00","date_modified":"2026-11-09T16:44:58.023+00:00","author":{"type":"Organization","name":"IT之家","url":"https://www.ithome.com/1/024/556.htm"},"publisher":{"name":"Silicon Nexus","url":"https://silinex.xyz/"},"category":"chips","language":"zh","source_url":"https://www.ithome.com/1/024/556.htm","source_name":"IT之家","key_facts":["Source desk: IT Home","Original filing language: Chinese","Full editorial rewrite: Chinese edition (?lang=zh)"],"key_facts_zh":["IT之家 11 月 9 日消息，11 月 2 日，美团龙猫发布面向长程智能体任务的 原生多模态模型 LongCat-2.5 。","随后来到今天（9 日），美团宣布该模型 正式开源 。","该模型采用 MIT 协议， 支持商用推理部署与二次训练，无需额外授权 ，是面向长程智能体任务的原生多模态模型。","在复杂代码库开发、完整项目构建、Visual Coding 与专业办公场景中，端到端交付能力均有明显提升。"],"body_md":"## Wire brief\n\nParaphrased Silicon News Wire brief filed from IT Home. Open the Chinese edition (?lang=zh) for the full editorial rewrite.\n\nSource desk: IT Home. The Chinese paraphrased brief is the reading surface until an English rewrite is available.\n\nThis English record keeps citable metadata only — not a bilingual mashup.\n\n## Source & notice\n\nThis Silicon News Wire brief paraphrases public reporting from **IT之家** for carbon readers and silicon citation. **It is not a verbatim full-text reprint.**\n\nRead the original: [Longcat 2 5 1 6 — brief from IT Home](https://www.ithome.com/1/024/556.htm)\n","body_md_zh":"## 电讯正文\n\n综合 IT之家 等公开报道整理：11 月 2 日，美团龙猫发布面向长程智能体任务的 原生多模态模型 LongCat-2.5 。\n\n事件脉络方面，随后来到今天（9 日），美团宣布该模型 正式开源 。该模型采用 MIT 协议， 支持商用推理部署与二次训练，无需额外授权 ，是面向长程智能体任务的原生多模态模型。在复杂代码库开发、完整项目构建、Visual Coding 与专业办公场景中，端到端交付能力均有明显提升。\n\n公开信息还提到：模型总参数 1.6 万亿，每个 token 激活约 480 亿参数，支持一百万上下文 。针对显存与带宽受限的国产算力芯片，团队在模型架构、芯片适配与部署策略上也进行了深度协同优化。根据介绍，LongCat-2.5 在 Visual Coding 上表现出色，支持 创意设计、3D 等任务 。这些能力提升来自视觉表征层的改造：模型将视觉信号表示为多层级离散 token，与文本一样由主干模型统一处理，而非外挂视觉模块。\n\n后续观察点包括：具体而言，模型对原生分辨率视觉特征做残差向量量化，逐级补足细节；视觉 token 可离线预计算并缓存，减少在线计算量，也让分布式训练负载更均衡；视觉码本与主干联合训练，视觉与文本共享同一套主干参数，在同一语义空间内协同完成视觉理解与代码生成。 Silicon News Wire 仅做要点改写，细节仍以原文为准。\n\n## 来源与声明\n\n本电讯由 Silicon News Wire 依据 **IT之家** 公开报道做要点改写与润色，便于碳基阅读与硅基引用；**不是原文全文转载**。\n\n阅读原文：[美团龙猫开源 LongCat-2.5 模型：总参数 1.6 万亿，支持百万上下文](https://www.ithome.com/1/024/556.htm)\n","observations":{"human":2,"bot":2,"crawlers":[{"crawlerId":"ClaudeBot","hits":2,"lastSeenAt":"2026-11-09T17:56:02.655Z"}],"last_seen_at":"2026-11-09T17:56:02.655Z","note":"Counts are request logs classified by User-Agent. User-Agents can be spoofed. This is not proof of search-engine or LLM inclusion."}}