{
 "schema_version": "1.0",
 "dataset": "deepseek-industry-applications",
 "verified_at": "2026-11-05",
 "note": "Counts of DeepSeek models/projects/plugins per industry, matched by keyword on descriptions. Not a complete census — a starting point.",
 "industries": [
  {
   "slug": "medical",
   "name_zh": "医疗",
   "name_en": "Medical",
   "models": 189,
   "projects": 16,
   "plugins": 242,
   "papers": 6,
   "notes": [
    "医疗场景对幻觉零容忍。模型能写病历摘要，不能做诊断结论。",
    "国内医疗 AI 需二类/三类器械审批，微调模型不等于可上线的医疗产品。",
    "患者数据上云前先看《个人信息保护法》和医院的数据出境规定。"
   ]
  },
  {
   "slug": "finance",
   "name_zh": "金融",
   "name_en": "Finance",
   "models": 86,
   "projects": 31,
   "plugins": 310,
   "papers": 6,
   "notes": [
    "金融输出必须可追溯。DeepSeek 的推理链能看，但不等于合规留痕。",
    "行情类任务对延迟敏感，先用离峰 API 省钱，但要注意离峰时段和交易时段错开。",
    "投顾、荐股类内容在国内需要持牌，模型输出不能直接对客。"
   ]
  },
  {
   "slug": "legal",
   "name_zh": "法律",
   "name_en": "Legal",
   "models": 80,
   "projects": 37,
   "plugins": 116,
   "papers": 6,
   "notes": [
    "法条会变。模型训练数据有截止日期，引用具体条款前必须人工核验现行版本。",
    "合同审查适合做「初筛」不适合做「定稿」，责任仍在人。",
    "涉外业务注意数据出境；案卷材料不要喂给公有云 API。"
   ]
  },
  {
   "slug": "education",
   "name_zh": "教育",
   "name_en": "Education",
   "models": 177,
   "projects": 16,
   "plugins": 68,
   "papers": 6,
   "notes": [
    "教育内容容错率低，数学/理科答案要验证，文科答案要人工复核。",
    "面向未成年人的产品有额外合规要求（内容审核、防沉迷）。",
    "题库类应用注意版权，模型生成的题目不能直接当作原创。"
   ]
  },
  {
   "slug": "manufacturing",
   "name_zh": "制造",
   "name_en": "Manufacturing",
   "models": 35,
   "projects": 8,
   "plugins": 36,
   "papers": 6,
   "notes": [
    "工业场景多数需要私有化部署，公有云 API 在数据安全评审时通常过不了。",
    "产线对延迟和稳定性要求高，7x24 可用性比模型能力更重要。",
    "设备手册、工艺参数是核心资产，落地前先想清楚数据边界。"
   ]
  },
  {
   "slug": "ecommerce",
   "name_zh": "电商",
   "name_en": "E-commerce",
   "models": 0,
   "projects": 16,
   "plugins": 42,
   "papers": 2,
   "notes": [
    "商品文案生成是最容易见效的场景，但要注意平台对 AI 生成内容的标注要求。",
    "客服机器人要设置兜底转人工，模型答错商品参数会直接产生售后成本。",
    "评价分析、竞品监控类应用注意爬虫合规。"
   ]
  }
 ]
}