{
	"nodes":[
		{"id":"group_models","type":"group","x":380,"y":-460,"width":480,"height":700,"color":"2","label":"二、模型演进历程"},
		{"id":"group_tech","type":"group","x":1020,"y":-460,"width":450,"height":630,"color":"5","label":"三、核心技术创新"},
		{"id":"group_competition","type":"group","x":880,"y":1240,"width":450,"height":620,"color":"3","label":"六、市场竞争定位"},
		{"id":"group_strategy","type":"group","x":80,"y":1280,"width":480,"height":560,"color":"6","label":"五、差异化竞争策略"},
		{"id":"group_cost","type":"group","x":-720,"y":480,"width":460,"height":560,"color":"1","label":"四、成本优势分析"},
		{"id":"group_company","type":"group","x":-200,"y":-300,"width":460,"height":520,"color":"4","label":"一、公司背景与团队"},
		{"id":"node_founder","type":"text","text":"## 创始人：梁文锋\n\n- 📅 1985年生，广东人\n- 🎓 浙江大学毕业\n- 💼 幻方量化创始人\n- 🏆 中国顶级量化投资专家","x":-180,"y":-250,"width":410,"height":120,"color":"4"},
		{"id":"node_parent","type":"text","text":"## 母公司：幻方量化\n\n**High-Flyer Quantitative**\n\n- 💰 投资 $5亿+ 硬件设施\n- 🖥️ 10000+ GPU 集群\n- 📈 中国顶级量化基金\n- 🎯 为 AI 研发提供强力支撑","x":-180,"y":-110,"width":410,"height":130,"color":"4"},
		{"id":"node_team","type":"text","text":"## 团队构成\n\n- 👥 ~90人团队\n- 🇨🇳 纯中国工程师\n- 🎓 应届生+年轻开发者\n- 📊 平均年龄约25岁\n- 🏫 清华、浙大等名校","x":-180,"y":40,"width":200,"height":140,"color":"4"},
		{"id":"node_founding","type":"text","text":"## 成立时间\n\n📅 **2023年7月17日**\n\n不到2年快速崛起","x":40,"y":60,"width":190,"height":80,"color":"4"},
		{"id":"0f0fbe2cc23bc5eb","type":"file","file":"90 Resources/附件/Gemini_2026_01_14__2324.md","x":520,"y":-20,"width":400,"height":400},
		{"id":"a1b2c3d4e5f6a7b8","type":"text","text":"# DeepSeek 深度求索的崛起之路\n\n## 从 $600万 到挑战 GPT-5 的技术突破\n\n**核心亮点**:\n- 🚀 $600万训练成本（节省90%）\n- 🤖 671B参数，37B激活的MoE架构\n- 🌟 MIT开源策略，技术自主可控\n- 📊 MMLU 66分，代码生成接近GPT-5","x":400,"y":300,"width":500,"height":180,"color":"6"},
		{"id":"node_v3","type":"text","text":"## DeepSeek V3 (2024年12月)\n\n- 📦 **671B** 总参数，**37B** 激活\n- 💰 **$600万**训练成本\n- 🏆 性能大幅跃升\n- 🌟 256个专家模块\n- 🎯 每次激活8个专家","x":400,"y":-140,"width":440,"height":140,"color":"2"},
		{"id":"node_r1","type":"text","text":"## DeepSeek R1 (2025年1月)\n\n- 🧠 **推理模型**\n- 🎯 对标 OpenAI o1\n- 🔓 MIT 开源","x":400,"y":20,"width":440,"height":100,"color":"2"},
		{"id":"node_v1","type":"text","text":"## DeepSeek V1 (2023年12月)\n\n- 📦 **7B / 67B** 参数\n- 🎉 首次发布\n- 🔓 MIT 开源\n- 💡 初步验证技术路线","x":400,"y":-410,"width":440,"height":110,"color":"2"},
		{"id":"node_v2","type":"text","text":"## DeepSeek V2 (2024年)\n\n- 📦 **236B** 参数\n- 🧩 **MoE 架构**引入\n- 🔬 **MLA 技术**首次应用\n- ⚡ 推理效率大幅提升","x":400,"y":-280,"width":440,"height":120,"color":"2"},
		{"id":"node_moe","type":"text","text":"## MoE 混合专家架构\n\n- 📦 **671B** 总参数\n- ⚡ **37B** 激活参数\n- 🧩 **256** 个专家模块\n- 🎯 每次激活 **8** 个专家\n- 💡 稀疏激活降低成本","x":1040,"y":-410,"width":410,"height":140,"color":"5"},
		{"id":"node_mla","type":"text","text":"## MLA 多头潜在注意力\n\n- 📉 KV Cache 减少 **93.3%**\n- 💾 显存占用大幅降低\n- ⚡ 推理速度显著提升\n- 🎯 长文本处理更高效","x":1040,"y":-250,"width":410,"height":130,"color":"5"},
		{"id":"node_dsa","type":"text","text":"## ⭐ DSA 稀疏注意力\n\n- 🚀 **O(L²) → O(L×k)**\n- ⚡ 计算量削减 **84%**\n- 📊 长文本处理优势\n- 💎 V3.2 核心创新\n- 🏆 业界领先技术","x":1040,"y":-100,"width":410,"height":150,"color":"5"},
		{"id":"node_achievement","type":"text","text":"# 🏆 关键成果\n\n## 性能指标\n- 📊 **MMLU**: 66分\n- 💻 **代码生成**: 接近GPT-5\n- 🧮 **数学推理**: 超越多数闭源\n- 📝 **长文本**: 128K上下文\n\n## 市场影响\n- 🌐 **API调用量**: 全球前列\n- ⭐ **GitHub**: 8万+ stars\n- 👥 **社区活跃**: 中国AI标杆\n- 🔓 **开源贡献**: MIT许可\n\n## 技术突破\n- 💰 **成本**: 90%削减\n- ⚡ **效率**: 84%计算优化\n- 🚀 **速度**: 快速迭代","x":1600,"y":-360,"width":420,"height":320,"color":"6"},
		{"id":"node_trend","type":"text","text":"# 🔮 技术趋势洞察\n\n## 开源成为主流\n- 🔓 MIT许可降低使用门槛\n- 👥 社区驱动加速创新\n- 🌍 全球开发者共建生态\n\n## 成本优化是竞争焦点\n- 💰 $600万 vs $6000万\n- ⚡ 推理成本降低89%\n- 🎯 普惠AI成为可能\n\n## MoE架构成熟\n- 🧩 稀疏激活降低成本\n- 📈 性能与效率兼得\n- 🚀 成为主流架构\n\n## 中国AI崛起\n- 🇨🇳 技术自主可控\n- 💻 国产芯片适配\n- 🏆 国际竞争力提升","x":1660,"y":60,"width":420,"height":350,"color":"5"},
		{"id":"node_mtp","type":"text","text":"## MTP 多 Token 预测\n\n- 📈 提高训练效率\n- 🔢 **FP8** 混合精度\n- ⚡ 优化收敛速度\n- 💰 降低训练成本","x":1060,"y":320,"width":410,"height":130,"color":"5"},
		{"id":"node_v32","type":"text","text":"## ⭐ DeepSeek V3.2 (2025年12月)\n\n- 🚀 **DSA 稀疏注意力**\n- ⚡ **84%** 计算量削减\n- 🏆 当前最强版本\n- 📊 O(L²) → O(L×k)\n- 💎 长文本处理优势显著","x":560,"y":640,"width":440,"height":130,"color":"2"},
		{"id":"node_sources","type":"text","text":"# 📚 数据来源\n\n**调研时间**: 2026年1月\n\n**主要来源**:\n- 📄 DeepSeek 官网技术报告\n- 🤖 机器之心、量子位等科技媒体\n- 💻 GitHub 开源社区\n- 📊 各大模型评测平台\n- 📈 行业分析报告\n- 🌐 多轮 WebSearch 深度调研\n\n**生成工具**: Claude Code + research-to-diagram skill","x":1560,"y":1000,"width":420,"height":250,"color":"4"},
		{"id":"node_intl_comp","type":"text","text":"## 国际竞品\n\n- 🏆 **GPT-5.2** (73分)\n- 💻 **Claude 4.5** (编程标准)\n- 🌐 **Gemini 3 Pro** (多模态)\n- 🤖 **Grok 4.1** (推理)\n- 💰 价格昂贵，闭源","x":900,"y":1290,"width":410,"height":140,"color":"3"},
		{"id":"node_domestic_comp","type":"text","text":"## 国产竞品\n\n- 🌙 **Kimi K2** (67分，长文本)\n- 🧠 **GLM-4.7** (68分，智谱AI)\n- 🐼 **百川、文心**等\n- 💰 成本较高，部分闭源","x":900,"y":1450,"width":410,"height":140,"color":"3"},
		{"id":"node_position","type":"text","text":"## ⭐ DeepSeek 定位\n\n- 📊 **66分**综合评分\n- 💰 **低成本高性能**\n- 🔓 **开源社区标杆**\n- 🇨🇳 **技术自主可控**\n- 🎯 **垂直领域优势**\n- 💎 **性价比之王**\n- 🚀 **快速迭代创新**","x":900,"y":1610,"width":410,"height":220,"color":"3"},
		{"id":"node_opensource","type":"text","text":"## ⭐ 全栈开源策略\n\n- 🔓 **MIT 许可证**\n- 📦 模型权重完全公开\n- 📄 技术报告详尽透明\n- 👥 社区驱动发展\n- 🌍 全球开发者可用\n- 🎯 降低使用门槛","x":100,"y":1330,"width":440,"height":160,"color":"6"},
		{"id":"node_vertical","type":"text","text":"## 垂直领域深耕\n\n- 💻 **代码生成**专长\n- 🧮 **数学推理**优势\n- 📝 **长文本处理**能力\n- 🇨🇳 **中文性能**强劲\n- 🎯 专注核心场景","x":100,"y":1510,"width":440,"height":150,"color":"6"},
		{"id":"node_chip","type":"text","text":"## 中国芯片适配\n\n- 🇨🇳 **H800** 优化\n- 💻 国产算力支持\n- 🔐 技术自主可控\n- 🎯 摆脱依赖","x":100,"y":1680,"width":440,"height":120,"color":"6"},
		{"id":"node_training_cost","type":"text","text":"## 🏆 GPU 预训练成本\n\n### DeepSeek V3\n💰 **$600万** (¥4300万)\n\n### vs GPT-4o\n💸 **$5000-6000万**\n\n### ⭐ 节省 **90%**","x":-700,"y":530,"width":410,"height":170,"color":"1"},
		{"id":"node_inference_cost","type":"text","text":"## API 推理成本\n\n### DeepSeek\n- 输入: **$0.27**/百万 tokens\n- 输出: **$1.1**/百万 tokens\n\n### GPT-4o\n- 输入: **$2.5**/百万 tokens\n- 输出: **$10**/百万 tokens\n\n### ⭐ 节省 **89%**","x":-700,"y":720,"width":410,"height":180,"color":"1"},
		{"id":"node_hardware","type":"text","text":"## 硬件利用率优化\n\n- 🇨🇳 **H800** 芯片适配\n- 💻 国产算力支持\n- ⚡ 高效训练流程","x":-700,"y":920,"width":410,"height":100,"color":"1"},
		{"id":"2a85db81b2c96162","type":"file","file":"90 Resources/附件/PDF++/DeepSeek的崛起之路.pdf","x":560,"y":840,"width":400,"height":280}
	],
	"edges":[
		{"id":"edge_main_company","fromNode":"a1b2c3d4e5f6a7b8","fromSide":"left","toNode":"group_company","toSide":"right","color":"4","label":"背景"},
		{"id":"edge_main_models","fromNode":"a1b2c3d4e5f6a7b8","fromSide":"top","toNode":"group_models","toSide":"bottom","color":"2","label":"演进"},
		{"id":"edge_main_tech","fromNode":"a1b2c3d4e5f6a7b8","fromSide":"right","toNode":"group_tech","toSide":"left","color":"5","label":"创新"},
		{"id":"edge_main_cost","fromNode":"a1b2c3d4e5f6a7b8","fromSide":"bottom","toNode":"group_cost","toSide":"top","color":"1","label":"成本"},
		{"id":"edge_main_strategy","fromNode":"a1b2c3d4e5f6a7b8","fromSide":"bottom","toNode":"group_strategy","toSide":"top","color":"6","label":"策略"},
		{"id":"edge_main_comp","fromNode":"a1b2c3d4e5f6a7b8","fromSide":"right","toNode":"group_competition","toSide":"left","color":"3","label":"定位"},
		{"id":"edge_parent_v3","fromNode":"node_parent","fromSide":"right","toNode":"node_v3","toSide":"left","label":"资金支持"},
		{"id":"edge_team_v32","fromNode":"node_team","fromSide":"right","toNode":"node_v32","toSide":"left","color":"4","label":"研发"},
		{"id":"edge_v1_v2","fromNode":"node_v1","fromSide":"bottom","toNode":"node_v2","toSide":"top","label":"升级"},
		{"id":"edge_v2_v3","fromNode":"node_v2","fromSide":"bottom","toNode":"node_v3","toSide":"top","label":"突破"},
		{"id":"edge_v3_r1","fromNode":"node_v3","fromSide":"bottom","toNode":"node_r1","toSide":"top","label":"分支"},
		{"id":"edge_v3_v32","fromNode":"node_v3","fromSide":"bottom","toNode":"node_v32","toSide":"top","color":"1","label":"进化"},
		{"id":"edge_moe_v3","fromNode":"node_moe","fromSide":"left","toNode":"node_v3","toSide":"right","color":"5","label":"架构"},
		{"id":"edge_mla_v3","fromNode":"node_mla","fromSide":"left","toNode":"node_v3","toSide":"right","color":"5","label":"优化"},
		{"id":"edge_dsa_v32","fromNode":"node_dsa","fromSide":"left","toNode":"node_v32","toSide":"right","color":"1","label":"核心技术"},
		{"id":"edge_mla_cost","fromNode":"node_mla","fromSide":"bottom","toNode":"node_training_cost","toSide":"top","color":"5","label":"降低训练成本"},
		{"id":"edge_dsa_inference","fromNode":"node_dsa","fromSide":"bottom","toNode":"node_inference_cost","toSide":"top","color":"1","label":"降低推理成本"},
		{"id":"edge_cost_opensource","fromNode":"node_training_cost","fromSide":"right","toNode":"node_opensource","toSide":"left","color":"1","label":"低成本支撑开源"},
		{"id":"edge_opensource_position","fromNode":"node_opensource","fromSide":"right","toNode":"node_position","toSide":"left","color":"6","label":"差异化优势"},
		{"id":"edge_v32_achievement","fromNode":"node_v32","fromSide":"right","toNode":"node_achievement","toSide":"left","label":"成果"},
		{"id":"edge_main_trend","fromNode":"a1b2c3d4e5f6a7b8","fromSide":"right","toNode":"node_trend","toSide":"left","label":"趋势"}
	],
	"metadata":{
		"version":"1.0-1.0",
		"frontmatter":{}
	}
}