{
 "generated": "2026-11-02",
 "node": {
  "id": "model:sensenova-loop-9b",
  "type": "Model",
  "name": "SenseNova Loop-9B-Preview",
  "slug": "sensenova-loop-9b",
  "date": "2026-10-30",
  "category": "open-llm",
  "category_zh": "开放权重 LLM",
  "access": "open",
  "license": "Apache-2.0",
  "eps": [
   110
  ],
  "page": "/graph/models/#m-sensenova-loop-9b",
  "size": "约 9B 总参数（HF 张量合计 9.66B；元数据显示精度分布是 F32 9.41B、BF16 0.24B，和模型卡“BF16 主干、FP32 的 mHC 路由参数”的说法不一致，本刊没有逐文件核）的多模态循环（Looped）Transformer：32 层不同的解码层，其中第 4 至 27 层这一块共享权重、连跑两遍，每次前向做 56 层解码计算，模型卡给的“推理参数”约 14.6B；线性注意力与全注意力 3:1 交错；两路 mHC（流形约束超连接）残差；读文本和图像、输出文本；训练最长序列 262,144 token",
  "avail": "开放权重，不设门槛（HF 未设 gated）：HF、ModelScope、GitHub（OpenSenseNova/SenseNova-Loop）；要用模型卡指定的 Transformers 5.18.0 加自带代码加载；SGLang 和 vLLM 的支持写的是两个 PR（SGLang #47281、vLLM #64832），模型卡让用 PR 版本；模型卡写支持 MTP 加速，DSpark、DFlash2 草稿模型“在训练中”；技术报告标 TBD",
  "license_full": "Apache-2.0（HF 卡、GitHub LICENSE）",
  "facts": [
   "架构：把中间一段解码层改成参数共享的循环块，用多轮迭代加深有效计算深度，而不加参数；两路 mHC 管理循环段里的信息流，Sinkhorn 归一化迭代 20 次（模型卡）",
   "商汤称在 9B 参数下“媲美 27B”，并在 WorkBuddy Bench、RaccoonPPT Bench、AutomationBench 上超过 Qwen3.5-27B；消融里加 mHC 使 GPQA Diamond 提高 4.42 分、TeamBench 提高 13.34 分（厂商自述）。本刊没有找到第三方复现",
   "发布日出处：模型卡 News 写 “[2026.10.30] We release SenseNova-Loop-9B-Preview”；IT之家 11-01 稿写“商汤科技 10 月 30 日宣布开源”；雷峰网转载的商汤稿落款 10-31 10:51。HF 仓库 10-28 建、GitHub 首个提交 10-29，都不是发布日"
  ],
  "sources": [
   {
    "label": "HF 模型页（含模型卡）",
    "url": "https://huggingface.co/sensenova/SenseNova-Loop-9B-Preview",
    "kind": "official"
   },
   {
    "label": "GitHub",
    "url": "https://github.com/OpenSenseNova/SenseNova-Loop",
    "kind": "github"
   },
   {
    "label": "雷峰网转载的商汤稿",
    "url": "https://www.leiphone.com/category/industrynews/f9tAHG0kPopK1Nnw.html",
    "kind": "media"
   },
   {
    "label": "IT之家（新浪科技转载）：10 月 30 日宣布开源",
    "url": "https://finance.sina.com.cn/tech/digi/2026-11-01/doc-inixetzz6796313.shtml",
    "kind": "media"
   }
  ],
  "ep_tags": {
   "110": "边界日 · 10-30"
  },
  "developer_text": "商汤科技（SenseTime，OpenSenseNova）",
  "degree": 0.0011
 },
 "page": "https://xiaopingfeng.com/graph/e/model/sensenova-loop-9b/",
 "out": [
  {
   "type": "developed_by",
   "sentence": "SenseNova Loop-9B-Preview的开发方是商汤科技",
   "direction": "out",
   "id": "org:sensetime",
   "name": "商汤科技（SenseTime）",
   "unit": "",
   "role": "primary",
   "status": "base",
   "evidence": [
    {
     "kind": "ledger",
     "label": "本站新模型台账",
     "text": "新模型台账：开发者栏写作“商汤科技（SenseTime，OpenSenseNova）”；HF 模型页（含模型卡）",
     "url": "https://huggingface.co/sensenova/SenseNova-Loop-9B-Preview"
    }
   ]
  },
  {
   "type": "placed_in",
   "sentence": "SenseNova Loop-9B-Preview出现在 Landscape 的「开放权重语言模型」",
   "direction": "out",
   "id": "layer:ai-models--1",
   "name": "开放权重语言模型（含多模态）",
   "status": "derived",
   "evidence": [
    {
     "kind": "own",
     "label": "本站页面",
     "text": "AI 新模型台账：SenseNova Loop-9B-Preview（开放权重语言模型（含多模态））",
     "url": "https://xiaopingfeng.com/graph/models/#m-sensenova-loop-9b"
    }
   ]
  }
 ],
 "in": []
}