{
 "updated": "2026-11-03",
 "about": "端侧模型图谱：能在手机、电脑、车机、边缘设备本地跑的小模型和语音模型、推理框架和运行时、端侧 NPU 代表、设备端 AI 方案。每家写 type（创业 / 大厂产品线 / 开源组织 / 研究机构）、modality、代表产品、params 参数量、devices 支持设备、chips 支持芯片、formats 端侧格式、license 开源协议（kind：开源 / 开放权重 / 闭源）、benchmarks 基准数据，以及融资、估值和关键数据，每个数字附出处（src）。layer 对应 layers.json 的 code。product_line_of = 大公司的产品线。",
 "companies": [
  {
   "name": "Apple Intelligence",
   "layer": "a1",
   "url": "https://www.apple.com/apple-intelligence/",
   "hq": "库比蒂诺",
   "region": "海外",
   "type": "大厂产品线",
   "founded": 2024,
   "desc": "系统级端侧基础模型加私有云计算，Siri 与写作、修图等",
   "modality": "多模态",
   "products": [
    "AFM 3 Core",
    "AFM 3 Core Advanced",
    "Private Cloud Compute"
   ],
   "params": {
    "v": "AFM 3 Core 约 3B 稠密；AFM 3 Core Advanced 20B 稀疏，每次激活 1–4B，权重存于闪存",
    "src": "https://machinelearning.apple.com/research/introducing-third-generation-of-apple-foundation-models"
   },
   "devices": [
    "iPhone",
    "Mac"
   ],
   "chips": [
    "Apple ANE"
   ],
   "formats": [
    "Core ML"
   ],
   "product_line_of": "Apple",
   "public": true,
   "note": "第三代 AFM 与 Google 合作开发（Apple 官方）。中国大陆端侧服务「Apple智能」2026-07-08 备案、07-15 公示（网信办公告 https://www.shaanxijubao.cn/20260716/bd205b07aac632923be790bd38f29278.html ）。Apple 称详细技术报告将于夏末发布，第三方基准暂无。"
  },
  {
   "name": "Google Gemini Intelligence / AICore",
   "layer": "a1",
   "url": "https://developer.android.com/ai",
   "hq": "山景城",
   "region": "海外",
   "type": "大厂产品线",
   "founded": 2026,
   "desc": "Android 系统级端侧 AI：AICore 托管 Gemini Nano，Gemini Intelligence 做任务代办",
   "modality": "多模态",
   "products": [
    "Gemini Nano 4 Fast / Full",
    "AICore",
    "Gemini Intelligence"
   ],
   "params": {
    "v": "Gemini Nano 4 基于 Gemma 4（Fast 版基于 E2B，另有 E4B 档），支持 140+ 语言；Fast 版标 TPU",
    "src": "https://www.estadao.com.br/tecmundo/inteligencia-artificial/412164-google-lanca-gemini-nano-4-que-roda-ia-direto-no-celular/"
   },
   "devices": [
    "Android"
   ],
   "chips": [
    "Google Tensor TPU",
    "Qualcomm Hexagon",
    "Samsung Exynos NPU"
   ],
   "formats": [
    "LiteRT"
   ],
   "product_line_of": "Google",
   "public": true,
   "note": "2026-04-02 宣布 Nano 4 并开放 AICore 开发者预览；Gemini Intelligence 要求 12GB RAM，且至少 Nano 3（Trusted Reviews https://www.trustedreviews.com/news/samsungs-new-foldables-get-gemini-intelligence-before-pixels-do ；i10x 转述文档 https://i10x.ai/news/gemini-nano-4-android-ai-update ）。founded 取 Gemini Intelligence 首发年，Nano 首发更早，未核。"
  },
  {
   "name": "Microsoft Copilot+ PC / Foundry Local",
   "layer": "a1",
   "url": "https://learn.microsoft.com/windows/ai/windows-ai-comparison",
   "hq": "雷德蒙德",
   "region": "海外",
   "type": "大厂产品线",
   "desc": "Windows 端侧 AI：NPU PC 上的 Phi Silica，Foundry Local 本地跑 20 多个开源模型",
   "modality": "通用",
   "products": [
    "Phi Silica",
    "Foundry Local",
    "Windows ML"
   ],
   "devices": [
    "Windows"
   ],
   "chips": [
    "Qualcomm Hexagon",
    "Intel NPU",
    "AMD XDNA"
   ],
   "formats": [
    "ONNX"
   ],
   "key_data": [
    {
     "k": "Foundry Local",
     "v": "20+ 开源 LLM 与语音模型，OpenAI 兼容 API，任意 Windows 硬件",
     "date": "2026-08",
     "src": "https://learn.microsoft.com/zh-cn/windows/ai/windows-ai-comparison"
    }
   ],
   "product_line_of": "Microsoft",
   "public": true,
   "note": "Phi Silica 的参数量、评测微软 platform card 未公开，未写 params。founded 未查。Phi Silica 需 Copilot+ PC，来源 https://learn.microsoft.com/en-au/windows/ai/windows-ai-comparison 。"
  },
  {
   "name": "Samsung Galaxy AI",
   "layer": "a1",
   "url": "https://www.samsung.com/global/galaxy/galaxy-ai/",
   "hq": "水原",
   "region": "海外",
   "type": "大厂产品线",
   "desc": "Galaxy 手机端侧 AI，Fold8/Flip8 首发 Gemini Nano 4",
   "modality": "多模态",
   "products": [
    "Galaxy AI",
    "Gemini Nano 4（Z Fold8 / Flip8）"
   ],
   "params": {
    "v": "端侧用 Google Gemini Nano 4（Fast 版基于 Gemma 4 E2B）；Galaxy S26 仍为 Nano 3",
    "src": "https://www.estadao.com.br/tecmundo/inteligencia-artificial/412164-google-lanca-gemini-nano-4-que-roda-ia-direto-no-celular/"
   },
   "devices": [
    "Android"
   ],
   "chips": [
    "Qualcomm Hexagon",
    "Samsung Exynos NPU"
   ],
   "product_line_of": "Samsung",
   "public": true,
   "note": "Z Flip8/Fold8/Fold8 Ultra 为首批 Nano 4 机型（2026-07-22 Unpacked，SamMobile https://www.sammobile.com/news/samsungs-newest-foldable-phones-use-googles-gemini-nano-4-on-device-ai-model/ ；S26 仍用 Nano 3 见 https://www.netzwelt.de/news/257922-google-bestaetigt-gemini-nano-4-selbst-samsung-galaxy-s26-ultra-geht-leer.html ）。大陆「盖乐世AI」2026-07 备案（同上网信办公告）。founded 未查。"
  },
  {
   "name": "OPPO AndesGPT 端侧",
   "layer": "a2",
   "url": "https://www.oppo.com/",
   "hq": "东莞",
   "region": "中国",
   "type": "大厂产品线",
   "founded": 2023,
   "desc": "安第斯大模型的端侧版本，驱动小布助手与通话摘要等",
   "modality": "语言",
   "products": [
    "AndesGPT",
    "小布助手"
   ],
   "params": {
    "v": "Find X7（2023-12）首搭 70 亿参数端侧版；全系另有 700 亿、1800 亿云侧档",
    "src": "https://www.qcong.com.cn/index/article/details/id/922.html"
   },
   "devices": [
    "Android"
   ],
   "chips": [
    "Qualcomm Hexagon",
    "MediaTek APU"
   ],
   "product_line_of": "OPPO",
   "note": "AndesGPT 2026-07-08 备案、07-15 公示（https://www.shaanxijubao.cn/20260716/bd205b07aac632923be790bd38f29278.html）。初稿 ColorOS 17「128K 上下文、内存占用降 48%」只见公众号，未采；当前端侧最新参数（后续 3B 版说法见 CSDN 汇总，低权威）未核一手。参数 src 为媒体报道，权威度偏低。"
  },
  {
   "name": "vivo 蓝心端侧大模型",
   "layer": "a2",
   "url": "https://www.vivo.com.cn/",
   "hq": "东莞",
   "region": "中国",
   "type": "大厂产品线",
   "desc": "蓝心大模型的端侧版本，3B 多模态推理模型支撑个人化智能",
   "modality": "多模态",
   "products": [
    "蓝心 3B 端侧多模态推理模型",
    "蓝心个人智能框架",
    "OriginOS 6"
   ],
   "params": {
    "v": "3B 端侧多模态推理大模型（2025-10 发布）",
    "src": "https://business.cctv.cn/2025/10/11/ARTI91SzY8EpOc5YVhSAf2p7251011.shtml"
   },
   "devices": [
    "Android"
   ],
   "chips": [
    "Qualcomm Hexagon",
    "MediaTek APU"
   ],
   "product_line_of": "vivo",
   "note": "「vivo蓝心端侧大模型」2026-07-08 备案、07-15 公示（https://www.shaanxijubao.cn/20260716/bd205b07aac632923be790bd38f29278.html）。初稿的「30B MoE 预研」未能在一手来源核实，未写；2026 开发者大会（原系统7）信息仅见聚合页，参数未查到。founded 未查。"
  },
  {
   "name": "努比亚 豆包手机（NaviX）",
   "layer": "a2",
   "url": "https://www.nubia.com/",
   "hq": "深圳",
   "region": "中国",
   "type": "大厂产品线",
   "founded": 2025,
   "desc": "努比亚与字节合作的 AI 智能体手机，豆包手机助手加端侧模型",
   "modality": "多模态",
   "products": [
    "努比亚 NaviX Ultra（豆包二代手机）",
    "豆包手机助手",
    "努比亚豆包手机大模型"
   ],
   "devices": [
    "Android"
   ],
   "product_line_of": "中兴通讯 / 字节跳动",
   "note": "「努比亚豆包手机大模型」备案单位为上海申启纪元智能终端（中兴旗下），2026-07-08 备案、07-15 公示（https://www.shaanxijubao.cn/20260716/bd205b07aac632923be790bd38f29278.html）；豆包手机助手由字节关联公司另行备案（搜狐 2026-11-02 https://www.sohu.com/a/1096742454_121347613 ）。NaviX Ultra 2026-07-16 发布（北京日报 http://xinwen.bjd.com.cn/content/s6a58b67de4b0e45f3fd4bd1e.html ）。一代 2025-12 工程机 3,499 元、3 万台售罄。端侧模型参数、芯片未查到。"
  },
  {
   "name": "华为 小艺（HarmonyOS 端侧）",
   "layer": "a2",
   "url": "https://consumer.huawei.com/cn/",
   "hq": "深圳",
   "region": "中国",
   "type": "大厂产品线",
   "desc": "小艺助手的本地模型，跑通话摘要、文本润色与图库搜索等",
   "modality": "多模态",
   "products": [
    "小艺",
    "本地大模型",
    "HarmonyOS 7"
   ],
   "params": {
    "v": "原生 30B 入端模型（激活约 2B），常驻内存（媒体转述 HDC 2026 说法）",
    "src": "http://www.c114.net.cn/industry/92365.html"
   },
   "devices": [
    "Android"
   ],
   "chips": [
    "华为 麒麟 NPU"
   ],
   "product_line_of": "华为",
   "note": "「华为小艺AI大模型」2026-07-08 备案、07-15 公示（https://www.shaanxijubao.cn/20260716/bd205b07aac632923be790bd38f29278.html）。HarmonyOS 7.0.0.102 起端侧机型可查看本地模型版本与参数（IT之家 https://www.ithome.com/0/990/776.htm ，未给具体数字）。另有公众号称端侧盘古约 100 亿参数，与上述 30B（2B 激活）口径冲突，未采；参数 src 为媒体转述，需华为官方稿复核。founded 未查。"
  },
  {
   "name": "小米 澎湃 AI（MiMo 端侧）",
   "layer": "a2",
   "url": "https://www.mi.com/",
   "hq": "北京",
   "region": "中国",
   "type": "大厂产品线",
   "desc": "MiMo 端侧模型驱动超级小爱 2.0 与澎湃 OS 本地推理",
   "modality": "多模态",
   "products": [
    "MiMo 端侧模型",
    "超级小爱 2.0",
    "澎湃 OS 4"
   ],
   "params": {
    "v": "端侧模型覆盖 0.3B、1.3B、4B 等档（百度百科转述，完整列表被截断，需复核）",
    "src": "https://baike.baidu.com/item/%E5%B0%8F%E7%B1%B3%E5%A4%A7%E6%A8%A1%E5%9E%8B/69005522"
   },
   "devices": [
    "Android",
    "车机"
   ],
   "chips": [
    "Qualcomm Hexagon"
   ],
   "product_line_of": "小米",
   "note": "「小米澎湃AI」2026-07-08 备案、07-15 公示（https://www.shaanxijubao.cn/20260716/bd205b07aac632923be790bd38f29278.html）。小米18 Pro 系列 2026-09-23 发布，首发澎湃 OS 4 与 MiMo V2 端侧模型（凤凰周刊 http://www.phoenixweekly.com.cn/detil.php?id=28099 ）。云侧 MiMo-V2.6 不属本卡。参数 src 为百科，权威度低。founded 未查。"
  },
  {
   "name": "荣耀 MagicOS 端侧（魔法大模型）",
   "layer": "a2",
   "url": "https://www.honor.com/cn/",
   "hq": "深圳",
   "region": "中国",
   "type": "大厂产品线",
   "founded": 2024,
   "desc": "魔法大模型：荣耀自研端侧平台级 AI，驱动 YOYO 智能体",
   "modality": "多模态",
   "products": [
    "魔法大模型",
    "YOYO 智能体",
    "MagicOS 11"
   ],
   "params": {
    "v": "70 亿参数（2024-01 在 Magic6 首发）",
    "src": "https://www.honor.com/cn/news/magicos-launch-2024/"
   },
   "devices": [
    "Android"
   ],
   "chips": [
    "Qualcomm Hexagon"
   ],
   "product_line_of": "荣耀",
   "note": "不在 2026-07 首批 7 款手机端侧备案名单内（https://www.shaanxijubao.cn/20260716/bd205b07aac632923be790bd38f29278.html）；其魔法大模型 2024-04 通用大模型备案见百度百科（低权威）。MagicOS 9.0 起拆为语言/图像/语音/多模态四类，3B 语言模型说法仅见论坛，未采。2026-09 HGDC 发布的新一代 YOYO 智能体用 397B 参数模型，属云侧，不计入端侧。"
  },
  {
   "name": "AMD XDNA（Ryzen AI）",
   "layer": "h1",
   "url": "https://www.amd.com/",
   "hq": "圣克拉拉",
   "region": "海外",
   "type": "大厂产品线",
   "desc": "Ryzen AI 处理器内的 XDNA 2 NPU，面向 Copilot+ PC",
   "modality": "通用",
   "products": [
    "Ryzen AI 400 / PRO 400",
    "XDNA 2"
   ],
   "devices": [
    "Windows"
   ],
   "chips": [
    "AMD XDNA"
   ],
   "key_data": [
    {
     "k": "Ryzen AI 400 系列 NPU",
     "v": "XDNA 2 架构，最高 60 TOPS，Zen 5 CPU",
     "date": "2026-01",
     "src": "https://www.amd.com/zh-cn/newsroom/press-releases/2026-1-5-amd-expands-ai-leadership-across-client-graphics-.html"
    }
   ],
   "product_line_of": "AMD",
   "public": true,
   "note": "founded 未查。"
  },
  {
   "name": "Apple Neural Engine",
   "layer": "h1",
   "url": "https://www.apple.com/newsroom/2025/10/apple-unleashes-m5-the-next-big-leap-in-ai-performance-for-apple-silicon/",
   "hq": "库比蒂诺",
   "region": "海外",
   "type": "大厂产品线",
   "desc": "A/M 系列 SoC 内置神经网络引擎，跑 Apple Intelligence 端侧模型",
   "modality": "通用",
   "products": [
    "Neural Engine（M5 为 16 核）",
    "M5 GPU 内置 Neural Accelerator"
   ],
   "devices": [
    "iPhone",
    "Mac"
   ],
   "chips": [
    "Apple ANE"
   ],
   "formats": [
    "Core ML",
    "MLX"
   ],
   "key_data": [
    {
     "k": "M5 Neural Engine",
     "v": "16 核，台积电第三代 3nm；GPU 每核内置神经加速器，GPU 峰值 AI 算力较 M4 提升 4 倍以上",
     "date": "2025-10",
     "src": "https://images.apple.com/ae/newsroom/2025/10/apple-unleashes-m5-the-next-big-leap-in-ai-performance-for-apple-silicon/"
    }
   ],
   "product_line_of": "Apple",
   "public": true,
   "note": "Apple 官方不公布 Neural Engine 的 TOPS 数，未采媒体估算；founded 未查。"
  },
  {
   "name": "Google Tensor（TPU）",
   "layer": "h1",
   "url": "https://store.google.com/",
   "hq": "山景城",
   "region": "海外",
   "type": "大厂产品线",
   "desc": "Pixel 手机 Tensor SoC 内的 TPU，跑端侧 Gemini Nano",
   "modality": "通用",
   "products": [
    "Tensor G6",
    "Pixel 11 系列"
   ],
   "devices": [
    "Android"
   ],
   "chips": [
    "Google Tensor TPU"
   ],
   "key_data": [
    {
     "k": "Tensor G6",
     "v": "台积电增强版 3nm，TPU 算力较 G5 提升 50%，用于端侧 Gemini Nano",
     "date": "2026-08",
     "src": "https://www.gadgets360.com/mobiles/news/google-pixel-11-tensor-g6-chipset-3nm-tsmc-process-report-11908942"
    }
   ],
   "product_line_of": "Google",
   "public": true,
   "note": "节点以 Google 硬件 VP 在台北的确认为准（CNA 转述，经 Gadgets 360）；早先 2nm 传闻已被否定。founded 未查。"
  },
  {
   "name": "Intel NPU（Core Ultra）",
   "layer": "h1",
   "url": "https://www.intel.com/",
   "hq": "圣克拉拉",
   "region": "海外",
   "type": "大厂产品线",
   "desc": "Core Ultra 处理器内的 NPU，AI PC 端侧推理单元",
   "modality": "通用",
   "products": [
    "Core Ultra Series 3（Panther Lake）NPU 5"
   ],
   "devices": [
    "Windows"
   ],
   "chips": [
    "Intel NPU"
   ],
   "formats": [
    "ONNX"
   ],
   "key_data": [
    {
     "k": "Panther Lake 的 NPU 5",
     "v": "50 TOPS；整平台含 GPU 共 180 TOPS；CPU 用 Intel 18A",
     "date": "2026-01",
     "src": "https://thenextweb.com/news/intel-computex-2026-18a-panther-nova-lake-ai-pc"
    }
   ],
   "product_line_of": "Intel",
   "public": true,
   "note": "Intel Loihi 等存内研究在存算图谱，不在此卡。founded 未查。"
  },
  {
   "name": "MediaTek NPU（天玑）",
   "layer": "h1",
   "url": "https://www.mediatek.com/",
   "hq": "新竹（中国台湾）",
   "region": "海外",
   "type": "大厂产品线",
   "desc": "天玑旗舰 SoC 内的 NPU 990，另带常驻低功耗存算一体 NPU",
   "modality": "通用",
   "products": [
    "天玑 9500 / NPU 990",
    "生成式 AI 引擎 2.0"
   ],
   "devices": [
    "Android"
   ],
   "chips": [
    "MediaTek APU"
   ],
   "key_data": [
    {
     "k": "天玑 9500 的 NPU 990",
     "v": "峰值性能较前代提升 111%、峰值功耗降 56%；台积电第三代 3nm；支持 BitNet 1.58 低位宽模型",
     "date": "2025-09",
     "src": "https://tech.udn.com/tech/story/123151/9021762"
    }
   ],
   "customers": [
    "Samsung Galaxy Tab S12 系列（天玑 9500）"
   ],
   "product_line_of": "MediaTek",
   "public": true,
   "note": "Tab S12 采用见 CNA：https://www.cna.com.tw/news/afe/202610020275.aspx 。区域按两分法记海外，总部在中国台湾，请站点自行处理标注。TOPS 绝对值官方未给；founded 未查。"
  },
  {
   "name": "Qualcomm Hexagon NPU",
   "layer": "h1",
   "url": "https://www.qualcomm.com/",
   "hq": "圣迭戈",
   "region": "海外",
   "type": "大厂产品线",
   "desc": "骁龙手机与 PC 平台内的 Hexagon NPU，端侧大模型的主力算力",
   "modality": "通用",
   "products": [
    "Snapdragon 8 Elite Gen 5",
    "Snapdragon X2 Elite"
   ],
   "devices": [
    "Android",
    "Windows"
   ],
   "chips": [
    "Qualcomm Hexagon"
   ],
   "formats": [
    "QNN"
   ],
   "key_data": [
    {
     "k": "骁龙 8 Elite Gen 5 的 Hexagon NPU",
     "v": "较上代快 37%、每瓦性能高 16%，台积电 3nm",
     "date": "2025-09",
     "src": "https://9to5google.com/2025/09/24/snapdragon-8-elite-gen-5/"
    },
    {
     "k": "骁龙 X2 Elite 的 Hexagon NPU",
     "v": "80 TOPS，3nm",
     "date": "2025-09",
     "src": "https://www.jonpeddie.com/news/qualcomms-snapdragon-x2-elite-and-x2-elite-extreme/"
    }
   ],
   "customers": [
    "Samsung Galaxy Z Fold8（骁龙 8 Elite Gen 5 for Galaxy）"
   ],
   "product_line_of": "Qualcomm",
   "public": true,
   "note": "founded 未查。Fold8 搭载见 Trusted Reviews：https://www.trustedreviews.com/news/samsungs-new-foldables-get-gemini-intelligence-before-pixels-do 。收购 Nexa、Edge Impulse、Arduino 属 s 层，不在此卡。"
  },
  {
   "name": "Samsung Exynos NPU",
   "layer": "h1",
   "url": "https://semiconductor.samsung.com/",
   "hq": "水原",
   "region": "海外",
   "type": "大厂产品线",
   "desc": "Exynos 手机处理器内的 NPU，随 Exynos 2600 迈入 2nm",
   "modality": "通用",
   "products": [
    "Exynos 2600"
   ],
   "devices": [
    "Android"
   ],
   "chips": [
    "Samsung Exynos NPU"
   ],
   "key_data": [
    {
     "k": "Exynos 2600",
     "v": "业界首款 2nm GAA 手机 AP，32K MAC NPU，生成式 AI 性能较 Exynos 2500 提升 113%",
     "date": "2025-12",
     "src": "https://en.yna.co.kr/view/AEN20251219003600320"
    }
   ],
   "customers": [
    "Galaxy Z Flip8（Exynos 2600 + 12GB，运行 Gemini Nano 4）"
   ],
   "product_line_of": "Samsung",
   "public": true,
   "note": "Flip8 配置见 Trusted Reviews：https://www.trustedreviews.com/news/samsungs-new-foldables-get-gemini-intelligence-before-pixels-do 。三星 PIM/MRAM 在存算图谱，不在此卡。founded 未查。"
  },
  {
   "name": "华为 麒麟 NPU（达芬奇 / 昇腾）",
   "layer": "h1",
   "url": "https://www.hisilicon.com/",
   "hq": "深圳",
   "region": "中国",
   "type": "大厂产品线",
   "desc": "麒麟手机 SoC 内的 NPU，跑小艺端侧大模型",
   "modality": "通用",
   "products": [
    "麒麟 9030 / 9030 Pro"
   ],
   "devices": [
    "Android"
   ],
   "chips": [
    "华为 麒麟 NPU"
   ],
   "key_data": [
    {
     "k": "麒麟 9030 系列",
     "v": "华为称 NPU 算力较上代提升 70%（CPU +25%、GPU +40%）",
     "date": "2026-10",
     "src": "https://tech.ifeng.com/c/8xBTQRlDxOG"
    },
    {
     "k": "麒麟 9030 Pro 制程",
     "v": "中芯国际 N+3（SemiAnalysis 拆解），逻辑密度约 113 MTr/mm²，仅 DUV",
     "date": "2026-06",
     "src": "https://en.theblockbeats.news/flash/351280"
    }
   ],
   "product_line_of": "华为",
   "note": "NPU 的 TOPS 绝对值华为未公布，网传「>40 TOPS」仅见爆料站，未采。手机端侧模型见 a2 华为小艺。founded 未查。"
  },
  {
   "name": "Ambarella",
   "layer": "h2",
   "url": "https://www.ambarella.com/",
   "hq": "圣克拉拉",
   "region": "海外",
   "type": "大厂产品线",
   "founded": 2004,
   "desc": "视觉 SoC 内置 CVflow AI 加速器，覆盖摄像头与车载边缘推理",
   "modality": "视觉",
   "products": [
    "CV7",
    "CV3",
    "N1"
   ],
   "devices": [
    "车机",
    "Linux"
   ],
   "chips": [
    "Ambarella CVflow"
   ],
   "revenue": {
    "v": "$390.7M",
    "date": "FY2026（截至 2026-01-31）",
    "kind": "全年营收，同比 +37.2%，10-K 摘要转载",
    "src": "https://www.metatrader.com/ru/symbols/nasdaq/amba/documents/224624-annual-report-fy2026"
   },
   "key_data": [
    {
     "k": "边缘 AI 营收占比",
     "v": "约 80%（2026 财年 Q3）",
     "date": "2025-11",
     "src": "https://nz.finance.yahoo.com/news/ambarella-inc-amba-q3-2026-210123778.html"
    }
   ],
   "product_line_of": "Ambarella",
   "public": true,
   "also_in": [
    {
     "t": "AI 硬件图谱 · 端侧 AI 芯片",
     "u": "https://aiwatch-hardware-rankings.vercel.app/map/#g=x1"
    }
   ],
   "note": "NASDAQ: AMBA。它是上市公司，type 枚举无此项，记大厂产品线并置 pub。founded 2004 来自常识，未逐条核；营收引用页为 10-K 摘要镜像，建议复核 SEC 原文。"
  },
  {
   "name": "Axelera AI",
   "layer": "h2",
   "url": "https://axelera.ai",
   "hq": "埃因霍温",
   "region": "海外",
   "type": "创业",
   "founded": 2021,
   "desc": "数字存内计算加 RISC-V 的边缘推理芯片 Metis / Europa",
   "modality": "通用",
   "products": [
    "Metis",
    "Europa"
   ],
   "devices": [
    "Linux",
    "Windows"
   ],
   "chips": [
    "Axelera Metis",
    "Axelera Europa"
   ],
   "total_raised": "累计融资超 $4.5 亿（Reuters 2026-09 转述公司口径）",
   "funding": [
    {
     "date": "2026-02",
     "round": "未标轮次",
     "amount": "超 $250M",
     "lead": "Innovation Industries 领投（BlackRock、Samsung Catalyst Fund 跟投）",
     "src": "https://siliconangle.com/2026/02/24/edge-ai-chip-startup-axelera-ai-raises-250m-funding-round/"
    }
   ],
   "key_data": [
    {
     "k": "Metis 单芯片",
     "v": "214 TOPS，约 10W",
     "date": "2026-02",
     "src": "https://www.digitaltoday.co.kr/cn/view/28861/axelera-ai-raises-more-than-250-million-for-edge-ai-chips"
    },
    {
     "k": "Europa 单芯片",
     "v": "629 TOPS，每颗 45W",
     "date": "2026-09",
     "src": "https://www.storagereview.com/sv/news/axelera-europa-ships-629-tops-at-45w-per-aipu-in-validated-dell-xe5-and-supermicro-servers"
    },
    {
     "k": "客户数",
     "v": "600+ 客户（CEO 接受 Reuters 采访）",
     "date": "2026-09",
     "src": "https://www.investing.com/news/stock-market-news/european-chip-startup-axelera-wins-ai-factory-supply-deals-and-launches-second-chip-4901604"
    }
   ],
   "customers": [
    "Dell",
    "Supermicro",
    "Prodrive Technologies",
    "PwC Middle East"
   ],
   "also_in": [
    {
     "t": "存算一体芯片图谱",
     "u": "https://cim-chip-map.vercel.app/"
    },
    {
     "t": "AI 硬件图谱 · 端侧 AI 芯片",
     "u": "https://aiwatch-hardware-rankings.vercel.app/map/#g=x1"
    }
   ],
   "note": "与存算一体图谱重复（cim-chip-map，C 组），融资与算力两条沿用其 src；客户数初稿写 500，Reuters 最新为 600+，以后者为准。工艺节点、估值、营收未查到一手来源。"
  },
  {
   "name": "BrainChip",
   "layer": "h2",
   "url": "https://brainchip.com",
   "hq": "拉古纳希尔斯",
   "region": "海外",
   "type": "大厂产品线",
   "desc": "事件驱动脉冲神经网络芯片/IP Akida，边缘低功耗推理",
   "modality": "通用",
   "products": [
    "Akida",
    "AKD1500"
   ],
   "devices": [
    "MCU",
    "Linux"
   ],
   "chips": [
    "BrainChip Akida"
   ],
   "revenue": {
    "v": "US$1.89M，同比增长 374%",
    "date": "2025",
    "kind": "2025 年营收，年报，媒体转述",
    "src": "https://smallcaps.com.au/article/brainchip-licenses-akida-2-ip-to-edgeai-targeting-future-royalties"
   },
   "total_raised": "累计融资未汇总（上市公司多次配售）",
   "funding": [
    {
     "date": "2025-12",
     "round": "募资（机构配售）",
     "amount": "$25M（约 A$35M）",
     "lead": "未写明",
     "src": "https://www.design-reuse-embedded.com/news/202512068/brainchip-announces-25-million-usd-funding-ahead-of-ces-to-power-next-gen-edge-ai/"
    }
   ],
   "key_data": [
    {
     "k": "AKD1500",
     "v": "800 GOPS，低于 300 mW，22nm FD-SOI（GlobalFoundries），2026-06 量产出货",
     "date": "2026-06",
     "src": "https://brainchip.com/press/brainchip-announces-commercial-availability-and-production-shipments-of-akd1500-neuromorphic-processors/"
    }
   ],
   "customers": [
    "Parsons（Blue Ridge Envisioneering，首批 1 万颗订单）",
    "EDGEAI（Akida 2 IP 授权）",
    "GlobalFoundries（制造伙伴）"
   ],
   "product_line_of": "BrainChip",
   "public": true,
   "also_in": [
    {
     "t": "存算一体芯片图谱",
     "u": "https://cim-chip-map.vercel.app/"
    },
    {
     "t": "AI 硬件图谱 · 端侧 AI 芯片",
     "u": "https://aiwatch-hardware-rankings.vercel.app/map/#g=x1"
    }
   ],
   "note": "ASX: BRN，OTCQX: BRCHF。与存算一体图谱重复（cim 类脑组），数据沿用其 src。成立年份说法不一，省略；2026-09 财报会提良率低于预期仅见聚合页，未采。"
  },
  {
   "name": "DeepX",
   "layer": "h2",
   "url": "https://deepx.ai/",
   "hq": "首尔",
   "region": "海外",
   "type": "创业",
   "founded": 2018,
   "desc": "低功耗 NPU 芯片 DX-M1（三星 5nm 量产），布局机器人与工业边缘",
   "modality": "通用",
   "products": [
    "DX-M1",
    "DX-M2（2nm，计划 2027 年量产）"
   ],
   "devices": [
    "Linux"
   ],
   "chips": [
    "DeepX DX-M1",
    "DeepX DX-M2"
   ],
   "revenue": {
    "v": "约 33 亿韩元",
    "date": "2025",
    "kind": "全年营收，公司口径（韩媒转述）；2026-04 单月新订单即达此规模",
    "src": "https://www.asiae.co.kr/lang/print.htm?idxno=2026052916063021955&lang=en"
   },
   "total_raised": "累计融资未披露（公司称正洽谈本轮超 6,000 亿韩元，Reuters 2026-04；两口径不一致）",
   "funding": [
    {
     "date": "2026",
     "round": "D 轮（多次交割）",
     "amount": "前三次交割合计约 1,333 亿韩元，目标约 3,000 亿韩元",
     "lead": "新韩金融集团、BNW Investment 等",
     "src": "https://en.startuprecipe.co.kr/archives/29557"
    }
   ],
   "key_data": [
    {
     "k": "DX-M1",
     "v": "最高 25 TOPS，1–5W，三星 5nm，已量产",
     "date": "2026-05",
     "src": "https://www.asiae.co.kr/lang/print.htm?idxno=2026052916063021955&lang=en"
    },
    {
     "k": "DX-M2 目标",
     "v": "最高 80 TOPS，三星 2nm，计划明年量产",
     "date": "2026-05",
     "src": "https://www.asiae.co.kr/lang/print.htm?idxno=2026052916063021955&lang=en"
    }
   ],
   "customers": [
    "现代汽车集团（机器人计算平台合作）"
   ],
   "also_in": [
    {
     "t": "AI 硬件图谱 · 端侧 AI 芯片",
     "u": "https://aiwatch-hardware-rankings.vercel.app/map/#g=x1"
    }
   ],
   "note": "拟 IPO（先韩国），截至 2026-10 查到未上市。初稿「2026-08 D 轮、估值 $22 亿」来自 Caplight 估算，未采。总部城市初稿写城南（cim 同），此处写首尔，请复核。D 轮来源 startuprecipe 权威度低。Reuters：https://au.investing.com/news/stock-market-news/korean-ai-chip-startup-deepx-hyundai-work-on-robots-powered-by-generative-ai-4362708"
  },
  {
   "name": "EdgeCortix",
   "layer": "h2",
   "url": "https://www.edgecortix.com/",
   "hq": "东京",
   "region": "海外",
   "type": "创业",
   "founded": 2019,
   "desc": "SAKURA-II 边缘 AI 加速器，8W 内跑数十亿参数生成式模型",
   "modality": "通用",
   "products": [
    "SAKURA-II",
    "SAKURA-X（chiplet）"
   ],
   "devices": [
    "Linux"
   ],
   "chips": [
    "EdgeCortix SAKURA-II"
   ],
   "total_raised": "累计融资近 $1 亿（含股权与非稀释性政府补助，公司口径，2025-08）",
   "funding": [
    {
     "date": "2025-08",
     "round": "B 轮首次交割",
     "amount": "未披露（2024-12 起累计近 $1 亿，含政府补助）",
     "lead": "Yanmar Ventures、Pacific Bay Capital、NTT Finance 等",
     "src": "https://finbets.websol.barchart.com/?module=news&storyID=34213587&selected=news&symbol="
    },
    {
     "date": "2026-04",
     "round": "B 轮追加",
     "amount": "未披露",
     "lead": "MPower Partners、Axiro Semiconductor",
     "src": "https://thebridge.jp/2026/04/edgecortix-raises-funding-from-mpower-partners-and-axiro-semiconductor"
    }
   ],
   "key_data": [
    {
     "k": "SAKURA-II",
     "v": "60 TOPS（INT8）/ 30 TFLOPS（BF16），标准 8W",
     "date": "2026-04",
     "src": "https://thebridge.jp/2026/04/edgecortix-raises-funding-from-mpower-partners-and-axiro-semiconductor"
    }
   ],
   "also_in": [
    {
     "t": "AI 硬件图谱 · 端侧 AI 芯片",
     "u": "https://aiwatch-hardware-rankings.vercel.app/map/#g=x1"
    }
   ],
   "note": "初稿「超 ¥170 亿」出自 Dempa 剪报，未重新打开核对，未写。TDK Ventures 亦投 B 轮（barchart 另一稿）。"
  },
  {
   "name": "Hailo",
   "layer": "h2",
   "url": "https://hailo.ai/",
   "hq": "特拉维夫",
   "region": "海外",
   "type": "创业",
   "founded": 2017,
   "desc": "数据流架构边缘 AI 加速器，Hailo-10H 可本地跑 LLM/VLM",
   "modality": "通用",
   "products": [
    "Hailo-8",
    "Hailo-10H",
    "Hailo-15"
   ],
   "devices": [
    "Linux",
    "Windows"
   ],
   "chips": [
    "Hailo-8",
    "Hailo-10H"
   ],
   "total_raised": "累计融资约 $3.4 亿以上（2024 年 C 轮后，行业媒体口径；引用页为聚合页，权威度低）",
   "funding": [
    {
     "date": "2024",
     "round": "C 轮",
     "amount": "$120M",
     "lead": "未核",
     "src": "https://www.edgeir.com/tag/hailo"
    }
   ],
   "key_data": [
    {
     "k": "Hailo-10H",
     "v": "40 TOPS（INT4），典型功耗 2.5W，支持 LLM/VLM",
     "date": "2025-07",
     "src": "https://www.datacenterdynamics.com/en/news/hailo-releases-halio-10h-edge-ai-chip/"
    },
    {
     "k": "客户与开发者",
     "v": "100+ 客户，开发者社区超 1 万人",
     "date": "2026-09",
     "src": "https://ir.microchip.com/news-events/press-releases/detail/1415/microchip-technology-completes-acquisition-of-hailo"
    }
   ],
   "owner": "2026-07-24 签署收购协议，2026-09-21 完成：Microchip Technology 收购 Hailo（金额未披露，预计不对其财报有重大影响）",
   "also_in": [
    {
     "t": "AI 硬件图谱 · 端侧 AI 芯片",
     "u": "https://aiwatch-hardware-rankings.vercel.app/map/#g=x1"
    }
   ],
   "note": "owner 已核一手：Microchip 签约稿（经 Edge AI and Vision Alliance 转载）https://www.edge-ai-vision.com/category/provider/microchip-technology/ ；交割稿 https://ir.microchip.com/news-events/press-releases/detail/1415/microchip-technology-completes-acquisition-of-hailo （2026-09-21，GlobeNewswire）。初稿的 $4.5 亿（HardwareMap）与「裁员约一半」未核一手，未写；C 轮领投方、月份未核。"
  },
  {
   "name": "Kneron",
   "layer": "h2",
   "url": "https://www.kneron.com/",
   "hq": "圣迭戈",
   "region": "海外",
   "type": "创业",
   "founded": 2015,
   "desc": "可重构 NPU 边缘芯片，KL1140 面向端侧大模型",
   "modality": "通用",
   "products": [
    "KL730",
    "KL1140"
   ],
   "devices": [
    "Linux",
    "车机"
   ],
   "chips": [
    "Kneron KL730",
    "Kneron KL1140"
   ],
   "total_raised": "累计融资超 $2 亿（行业媒体口径，权威度低）",
   "funding": [
    {
     "date": "2023",
     "round": "未标轮次",
     "amount": "$4,900 万",
     "lead": "Foxconn Co-GP Fund 等",
     "src": "https://servernews.ru/1132890"
    }
   ],
   "key_data": [
    {
     "k": "KL730",
     "v": "最高 8 TOPS，支持 4K 60fps",
     "date": "2026-09",
     "src": "https://robottoday.com/suppliers-discovery/kneron-inc"
    }
   ],
   "also_in": [
    {
     "t": "AI 硬件图谱 · 端侧 AI 芯片",
     "u": "https://aiwatch-hardware-rankings.vercel.app/map/#g=x1"
    }
   ],
   "note": "KL1140「跑 1200 亿参数」属公司宣称，技术参数未公开，未写入数据。融资引用均为二手（SiliconAngle 经 ServerNews 转述），未找到一手稿，权威度低。"
  },
  {
   "name": "NXP eIQ Neutron NPU",
   "layer": "h2",
   "url": "https://www.nxp.com/",
   "hq": "埃因霍温",
   "region": "海外",
   "type": "大厂产品线",
   "desc": "i.MX 与 MCU 内的 eIQ Neutron NPU，面向工业与车载边缘",
   "modality": "通用",
   "products": [
    "i.MX 95",
    "eIQ Neutron NPU"
   ],
   "devices": [
    "Linux",
    "MCU",
    "车机"
   ],
   "chips": [
    "NXP eIQ Neutron"
   ],
   "key_data": [
    {
     "k": "i.MX 95",
     "v": "eIQ Neutron NPU 最高 8 eTOPS（NXP 的等效 TOPS 口径，非原始算力）",
     "date": "2026-04",
     "src": "https://www.mouser.com/en/new/nxp-semiconductors/nxp-toradex-next-gen-edge-devices-imx-9"
    }
   ],
   "product_line_of": "NXP",
   "public": true,
   "note": "NASDAQ: NXPI。TOPS 为 eTOPS 口径。founded 未查。"
  },
  {
   "name": "SiMa.ai",
   "layer": "h2",
   "url": "https://sima.ai/",
   "hq": "圣何塞",
   "region": "海外",
   "type": "创业",
   "founded": 2018,
   "desc": "Modalix 边缘 ML SoC 加 Palette 软件栈，面向机器人与无人机",
   "modality": "通用",
   "products": [
    "Modalix MLSoC",
    "Palette"
   ],
   "devices": [
    "Linux"
   ],
   "chips": [
    "SiMa.ai Modalix"
   ],
   "valuation": {
    "v": "$14.5 亿",
    "date": "2026-09",
    "kind": "C 轮估值，公司公布",
    "src": "https://siliconangle.com/2026/09/28/physical-ai-custom-chip-producer-sima-ai-raises-150m-at-1-45b-valuation/"
   },
   "total_raised": "累计融资约 $5 亿（公司口径）",
   "funding": [
    {
     "date": "2026-09",
     "round": "C 轮",
     "amount": "$150M",
     "lead": "Fidelity Management & Research、Amplify 联合领投（Dell Technologies Capital、Point72、AllianceBernstein、Baron Capital、J.P. Morgan 参与）",
     "post": "投后 $14.5 亿",
     "src": "https://siliconangle.com/2026/09/28/physical-ai-custom-chip-producer-sima-ai-raises-150m-at-1-45b-valuation/"
    }
   ],
   "also_in": [
    {
     "t": "AI 硬件图谱 · 端侧 AI 芯片",
     "u": "https://aiwatch-hardware-rankings.vercel.app/map/#g=x1"
    }
   ],
   "note": "计划推出 1,000 TOPS 级平台（Korthos 转述公司稿 https://korthosrobotics.com/research/sima-ai-raises-150m-physical-ai-compute ）；Modalix 的 TOPS 未查。"
  },
  {
   "name": "Syntiant",
   "layer": "h2",
   "url": "https://www.syntiant.com",
   "hq": "尔湾",
   "region": "海外",
   "type": "创业",
   "founded": 2017,
   "desc": "超低功耗神经决策处理器 NDP，常开语音与传感推理",
   "modality": "通用",
   "products": [
    "NDP120",
    "NDP200",
    "NDP250"
   ],
   "devices": [
    "MCU"
   ],
   "chips": [
    "Syntiant NDP"
   ],
   "revenue": {
    "v": "$271.8M",
    "date": "2025",
    "kind": "S-1 披露全年总营收，绝大部分来自 2024 年底收购的传感器业务，其中 AI 业务约 $12.3M",
    "src": "https://content-archive.fast-edgar.com/20260706/A82VQ22EZZ2RO9ZU22ZS2ZYLAKNMZZ22Z28Q/"
   },
   "total_raised": "累计融资超 $1.2 亿（OCBJ 2024-04；此后轮次未核一手）",
   "funding": [
    {
     "date": "2022-03",
     "round": "未标轮次",
     "amount": "$55M",
     "lead": "新进 Renesas、Millennium Technology Value Partners、Mirae Asset Capital 等",
     "src": "https://www.globenewswire.com/fr/news-release/2022/03/28/2411020/0/en/AI-Chip-Company-Syntiant-Raises-55-Million-to-Accelerate-Growth.html"
    }
   ],
   "key_data": [
    {
     "k": "部署规模",
     "v": "AI 技术部署于数千万台设备；SiSonic 传感器累计出货超 250 亿颗（截至 2026-03-31，S-1）",
     "date": "2026-03",
     "src": "https://stockanalysis.com/stocks/sytn/company/"
    },
    {
     "k": "2026Q1 营收",
     "v": "$6,450 万（去年同期 $6,660 万），净亏 $2,620 万",
     "date": "2026-03",
     "src": "https://thenextweb.com/news/syntiant-us-ipo-edge-ai-nasdaq"
    }
   ],
   "customers": [
    "Intel、Microsoft（M12）为投资方"
   ],
   "also_in": [
    {
     "t": "存算一体芯片图谱",
     "u": "https://cim-chip-map.vercel.app/"
    },
    {
     "t": "AI 硬件图谱 · 端侧 AI 芯片",
     "u": "https://aiwatch-hardware-rankings.vercel.app/map/#g=x1"
    }
   ],
   "note": "IPO 已核一手：2026-07-06 公司公告公开提交 S-1，拟纳斯达克上市，代码 SYTN，股数与价格未定（https://www.globenewswire.com/fr/news-release/2026/07/06/3322790/0/en/Syntiant-Publicly-Files-Registration-Statement-for-Proposed-Initial-Public-Offering.html ）；截至 2026-10-20 stockanalysis 仍显示未挂牌，故 pub 不填。与存算一体图谱重复（cim 近存组）。AI 业务金额出自中文转载的 S-1 表格，建议复核原文。"
  },
  {
   "name": "地平线 Horizon Robotics",
   "layer": "h2",
   "url": "https://www.horizon.auto/",
   "hq": "北京",
   "region": "中国",
   "type": "大厂产品线",
   "desc": "车载与机器人边缘 AI 芯片，征程 6 系列带 BPU 算力",
   "modality": "通用",
   "products": [
    "征程 6 系列",
    "Horizon SuperDrive"
   ],
   "devices": [
    "车机"
   ],
   "chips": [
    "Horizon BPU"
   ],
   "revenue": {
    "v": "37.6 亿元",
    "date": "2025",
    "kind": "全年营收，同比 +57.7%，毛利率 64.5%（港交所业绩公告）",
    "src": "https://stock.10jqka.com.cn/hks//20260323/c675465925.shtml"
   },
   "key_data": [
    {
     "k": "中高阶智驾方案出货",
     "v": "180 万套，同比近 5 倍",
     "date": "2025",
     "src": "http://yuanchuang.caijing.com.cn/2026/0325/5149532.shtml"
    }
   ],
   "product_line_of": "地平线",
   "public": true,
   "note": "港交所 9660.HK。以汽车为主（占营收 94.6%），端侧大模型相关度较低，收录为国内边缘 NPU 代表，可酌情删。founded 未查。"
  },
  {
   "name": "爱芯元智 Axera",
   "layer": "h2",
   "url": "https://www.axera-tech.com/",
   "hq": "宁波",
   "region": "中国",
   "type": "创业",
   "founded": 2019,
   "desc": "自研混合精度 NPU 加 AI-ISP 的边缘与端侧推理 SoC",
   "modality": "视觉",
   "products": [
    "AX650N",
    "AX630C"
   ],
   "devices": [
    "Linux",
    "车机"
   ],
   "chips": [
    "Axera NPU（爱芯通元）"
   ],
   "revenue": {
    "v": "4.73 亿元",
    "date": "2024",
    "kind": "全年营收，招股书；2022 年为 5,023 万元",
    "src": "https://www.yilantop.com/news/92890"
   },
   "total_raised": "累计融资未披露（已上市）",
   "funding": [
    {
     "date": "2026-02",
     "round": "港交所 IPO",
     "amount": "募资总额约 29.59 亿港元（净额约 27.99 亿港元）",
     "lead": "基石投资者含韦尔半导体香港等",
     "src": "http://news.cnfol.com/zhengquanyaowen/20260210/32010115.shtml"
    }
   ],
   "key_data": [
    {
     "k": "市占",
     "v": "按 2024 出货量，全球视觉端侧 AI 推理芯片第 5，份额 6.8%（灼识咨询）",
     "date": "2024",
     "src": "https://www.yilantop.com/news/92890"
    }
   ],
   "public": true,
   "note": "2026-02-10 在港交所挂牌（06809.HK），发行价 28.2 港元。总部宁波镇海（百度百科）与上海（部分报道）两种口径，取宁波。招股书称累计交付超 1.65 亿颗 SoC，见 https://www.iotworld.com.cn/html/News/202601/de5092e8ac3996b1.shtml 。"
  },
  {
   "name": "瑞芯微 Rockchip",
   "layer": "h2",
   "url": "https://www.rock-chips.com/",
   "hq": "福州",
   "region": "中国",
   "type": "大厂产品线",
   "desc": "RK3588 等 SoC 内置自研 NPU，RK182X 协处理器专跑端侧大模型",
   "modality": "通用",
   "products": [
    "RK3588",
    "RK3576",
    "RK182X"
   ],
   "devices": [
    "Linux",
    "Android",
    "车机"
   ],
   "chips": [
    "Rockchip NPU",
    "RK182X"
   ],
   "revenue": {
    "v": "44.02 亿元",
    "date": "2025",
    "kind": "全年营收，同比 +40.36%；净利 10.40 亿元",
    "src": "https://wallstreetcn.com/articles/3769961"
   },
   "key_data": [
    {
     "k": "RK182X",
     "v": "3D 架构端侧算力协处理器，2025 年 Q3 发布；2026 中报称数百个客户在研发，首批进入量产",
     "date": "2026-08",
     "src": "https://4g.stockstar.com/detail/RB2026081800002294"
    },
    {
     "k": "端侧 LLM 速度",
     "v": "SoC+RK182X 方案跑 3B/7B LLM 与多模态 VLM 近百 token/s（券商转述公司说法）",
     "date": "2026-04",
     "src": "https://dxpress.gelonghui.com/news/5217306"
    }
   ],
   "product_line_of": "瑞芯微",
   "public": true,
   "note": "上交所 603893。RK3588 的 6 TOPS 等规格本轮未重新核，未写；Moonshine 针对其 NPU 的说法见 https://faun.dev/company/moonshine-ai/ 。founded 未查。"
  },
  {
   "name": "Arm Ethos NPU",
   "layer": "h3",
   "url": "https://www.arm.com/products/silicon-ip-cpu/ethos/ethos-u85",
   "hq": "剑桥",
   "region": "海外",
   "type": "大厂产品线",
   "desc": "授权给 MCU 与嵌入式芯片的 Ethos-U 系列 NPU IP，U85 支持 Transformer",
   "modality": "通用",
   "products": [
    "Ethos-U85",
    "Ethos-U55 / U65"
   ],
   "devices": [
    "MCU",
    "Linux"
   ],
   "chips": [
    "Arm Ethos-U"
   ],
   "key_data": [
    {
     "k": "Ethos-U85",
     "v": "128–2048 MAC，1GHz 下最高 4 TOPS；原生支持 Transformer；能效较 U55/U65 高 20%",
     "date": "2025-03",
     "src": "https://www.arm.com/products/silicon-ip-cpu/ethos/ethos-u85"
    }
   ],
   "customers": [
    "NXP i.MX 93（Ethos-U65）"
   ],
   "product_line_of": "Arm",
   "public": true,
   "note": "i.MX 93 采用 Ethos-U65 见 https://emb-exp.com/nxp-npus-understanding-the-shift/（低权威）。founded 未查。"
  },
  {
   "name": "Expedera",
   "layer": "h3",
   "url": "https://www.expedera.com/",
   "hq": "圣克拉拉",
   "region": "海外",
   "type": "创业",
   "desc": "授权 Origin 系列 NPU IP，面向手机、车载与边缘",
   "modality": "通用",
   "products": [
    "Origin NPU IP"
   ],
   "devices": [
    "Android",
    "车机"
   ],
   "chips": [
    "Expedera Origin"
   ],
   "total_raised": "累计融资超 $4,700 万（公司口径，2024-05）",
   "funding": [
    {
     "date": "2024-05",
     "round": "B 轮",
     "amount": "$20M",
     "lead": "indie Semiconductor 领投",
     "src": "https://www.design-reuse-china.com/news/202405119/"
    }
   ],
   "customers": [
    "indie Semiconductor（战略投资，用于 ADAS 产品线）"
   ],
   "note": "founded、2024-05 之后的融资、公开 TOPS 数据未查到；估值未披露。"
  },
  {
   "name": "Quadric",
   "layer": "h3",
   "url": "https://quadric.io/",
   "hq": "伯林盖姆",
   "region": "海外",
   "type": "创业",
   "desc": "授权 Chimera 通用 NPU（GPNPU）IP，可跑视觉与端侧 LLM",
   "modality": "通用",
   "products": [
    "Chimera GPNPU"
   ],
   "devices": [
    "车机",
    "Windows"
   ],
   "chips": [
    "Quadric Chimera"
   ],
   "total_raised": "累计融资约 $9,000 万（FinSMEs 报道，2026-07；公司 2026-01 稿称 $7,200 万）",
   "funding": [
    {
     "date": "2026-01",
     "round": "C 轮",
     "amount": "$30M",
     "lead": "ACCELERATE Fund（BEENEXT 管理）领投",
     "src": "https://semiiphub.com/pulse/news/quadric-series-c-funding"
    },
    {
     "date": "2026-07",
     "round": "C 轮延长",
     "amount": "C 轮累计 $46M",
     "lead": "International Finance Corporation 领投",
     "src": "https://www.finsmes.com/2026/07/quadric-extends-series-c-to-46m.html"
    }
   ],
   "key_data": [
    {
     "k": "Chimera 算力",
     "v": "最高 864 TOPS，含车规配置；可部署最高 300 亿参数 LLM",
     "date": "2026-01",
     "src": "https://www.photonics.com/Articles/Quadric-Closes-30M-Series-C-Funding-Round/p21/a71945"
    },
    {
     "k": "产品营收",
     "v": "2025 年较 2024 年增长超 2 倍",
     "date": "2026-01",
     "src": "https://semiiphub.com/pulse/news/quadric-series-c-funding"
    }
   ],
   "note": "founded、估值、具名客户未查到；累计融资两个口径不同，已并列。"
  },
  {
   "name": "Ai2 OLMo",
   "layer": "m1",
   "url": "https://allenai.org/olmo",
   "hq": "西雅图",
   "region": "海外",
   "type": "研究机构",
   "founded": 2024,
   "founded_month": "2024-02",
   "desc": "全开放（权重、数据、代码）的 OLMo 系列，含 1B 与 7B 级",
   "modality": "语言",
   "products": [
    "OLMo 2 1B",
    "Olmo 3 7B"
   ],
   "params": {
    "v": "OLMo 2 1B；Olmo 3 7B（另有 13B、32B）",
    "src": "https://docs.allenai.org/models/olmo/index.html"
   },
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://www.aquanode.io/models/olmo"
   },
   "note": "Ai2 是非营利研究机构；OLMo 主打开放而非端侧，我们未查到官方的端侧设备或芯片信息，因此 devices/chips/formats 留空。Bolmo-1B（字节级，基于 OLMo 2 1B）有开放权重。基准未查到。端侧定位弱，可酌情去掉。"
  },
  {
   "name": "Apple Foundation Models",
   "layer": "m1",
   "url": "https://developer.apple.com/apple-intelligence/",
   "hq": "美国库比蒂诺",
   "region": "海外",
   "type": "大厂产品线",
   "founded": 2024,
   "founded_month": "2024-06",
   "desc": "iPhone、Mac 内置设备端基础模型，经 Foundation Models 框架开放给 App",
   "modality": "多模态",
   "products": [
    "AFM 3 Core",
    "AFM 3 Core Advanced（稀疏 20B）"
   ],
   "params": {
    "v": "端侧稠密约 3B（前代）；第三代 AFM 3 Core Advanced 为 20B 稀疏模型，每个 prompt 激活 1–4B",
    "src": "https://ofox.ai/zh/blog/apple-foundation-models-3-wwdc-2026-developer-jiedu/"
   },
   "devices": [
    "iPhone",
    "Mac"
   ],
   "chips": [
    "Apple ANE"
   ],
   "license": {
    "v": "闭源（仅经系统框架调用）",
    "kind": "闭源",
    "src": "https://machinelearning.apple.com/research/introducing-third-generation-of-apple-foundation-models"
   },
   "benchmarks": [
    {
     "k": "AFM 3 Core 人评偏好（对 2025 年基线，文本）",
     "v": "45.6% 偏好新模型（基线 23.3%）",
     "date": "2026-06",
     "src": "https://ofox.ai/blog/apple-foundation-models-3-wwdc-2026-developer-read/"
    }
   ],
   "product_line_of": "Apple",
   "public": true,
   "note": "第三代基础模型 2026-06-08 WWDC 发布，共五个模型（Gigazine、ofox）。Apple 只公布对自家 2025 基线的人评偏好，没有 MMLU 等公开基准（ofox 转述）。第三代部分与 Google 合作开发（Gigazine 报道）。端侧参数细节来自第三方解读，Apple 技术页面未逐字核。"
  },
  {
   "name": "Google Gemma / Gemini Nano",
   "layer": "m1",
   "url": "https://ai.google.dev/gemma",
   "hq": "美国山景城",
   "region": "海外",
   "type": "大厂产品线",
   "founded": 2024,
   "founded_month": "2024-02",
   "desc": "开放 Gemma 小模型，并作为系统内置 Gemini Nano 经 AICore 供 Android 本地调用",
   "modality": "多模态",
   "products": [
    "Gemma 4 E2B",
    "Gemma 4 E4B",
    "Gemini Nano 4（基于 Gemma 4）"
   ],
   "params": {
    "v": "Gemma 4 E2B 有效 2.3B（总 5.1B）/ E4B 有效 4.5B（总 8B）",
    "src": "https://wavespeed.ai/blog/posts/what-is-google-gemma-4/"
   },
   "devices": [
    "Android",
    "iPhone",
    "Mac",
    "Windows",
    "Linux",
    "树莓派",
    "浏览器"
   ],
   "chips": [
    "Qualcomm Hexagon",
    "MediaTek APU",
    "NVIDIA GPU"
   ],
   "formats": [
    "LiteRT",
    "GGUF",
    "MLX"
   ],
   "license": {
    "v": "Apache-2.0（Gemma 4）",
    "kind": "开源",
    "src": "https://developers.google.com/edge/litert-lm/models/gemma-4"
   },
   "key_data": [
    {
     "k": "Raspberry Pi 5 上 E2B 解码速度",
     "v": "7.6 tok/s（Google 报告，WaveSpeed 转述）",
     "date": "2026-04",
     "src": "https://wavespeed.ai/blog/posts/what-is-google-gemma-4/"
    }
   ],
   "product_line_of": "Google",
   "public": true,
   "note": "Gemma 4 于 2026-04-02 发布并改用 Apache 2.0（9to5Google）。Gemini Nano 4 基于 Gemma 4，2026 年内登陆高端 Android 机，Gemini Nano 本身不开放权重（闭源）。基准未查到：Artificial Analysis 页有 Gemma 4 E4B 条目但数值未核。设备/芯片列表中除 Qualcomm、MediaTek 合作（WaveSpeed 转述 Google 说法）外未逐项核。来源 9to5google https://9to5google.com/2026/04/02/google-gemma-4/ ；AICore https://android-dot-devsite-v2-prod.appspot.com/blog/posts/announcing-gemma-4-in-the-ai-core-developer-preview?hl=zh-tw"
  },
  {
   "name": "Hugging Face SmolLM",
   "layer": "m1",
   "url": "https://huggingface.co/HuggingFaceTB",
   "hq": "纽约",
   "region": "海外",
   "type": "开源组织",
   "founded": 2024,
   "founded_month": "2024-07",
   "desc": "完全开放（含数据与训练配方）的 SmolLM / SmolVLM 小模型",
   "modality": "多模态",
   "products": [
    "SmolLM3 3B",
    "SmolLM2 135M/360M/1.7B",
    "SmolVLM 256M"
   ],
   "params": {
    "v": "SmolLM3 3B；SmolLM2 135M / 360M / 1.7B；SmolVLM 最小 256M",
    "src": "https://nextomoro.com/hugging-face/"
   },
   "devices": [
    "iPhone",
    "Mac",
    "浏览器",
    "Android"
   ],
   "formats": [
    "GGUF",
    "LiteRT",
    "ONNX"
   ],
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://huggingface.co/jc-builds/SmolLM3-3B-Instruct-GGUF/blob/2851fb8db59cdcf47d832e14332e131e5e545c3d/README.md"
   },
   "valuation": {
    "v": "$45 亿",
    "date": "2023-08",
    "kind": "投后估值，媒体报道",
    "src": "https://thenextweb.ru/news/hugging-face-exploring-sale-13bn-valuation/"
   },
   "total_raised": "累计融资约 $4 亿（nextomoro 口径，低权威；D 轮后未再披露外部融资）",
   "funding": [
    {
     "date": "2021-03",
     "round": "B 轮",
     "amount": "$4000 万",
     "src": "https://nextomoro.com/hugging-face/"
    },
    {
     "date": "2022-05",
     "round": "C 轮",
     "amount": "$1 亿",
     "src": "https://nextomoro.com/hugging-face/"
    },
    {
     "date": "2023-08",
     "round": "D 轮",
     "amount": "$2.35 亿",
     "lead": "Salesforce Ventures（跟投 Google、Amazon、Nvidia、Intel、Qualcomm、IBM 等）",
     "post": "投后 $45 亿",
     "src": "https://thenextweb.ru/news/hugging-face-exploring-sale-13bn-valuation/"
    }
   ],
   "key_data": [
    {
     "k": "SmolLM3-3B 在 iPhone 17 Pro 上解码（LiteRT-LM int4，社区转换）",
     "v": "约 22.5 tok/s",
     "date": "2026-11",
     "src": "https://github.com/john-rocky/hf-to-litertlm/blob/97d1292f13c220834d689a3a5a766781a9782623/cards/smollm3-3b-litert.md"
    }
   ],
   "note": "Hugging Face 整体是模型托管平台，这里只收其 SmolLM/SmolVLM 小模型线。B、C 轮的金额与时间来自 nextomoro（低权威），领投方未查。Business Insider 经 TNW 俄文版转述：2026-08 公司聘银行家试探以 $130 亿以上估值出售（早期阶段，买方未定），未经公司确认。基准未查到（SmolLM3 官方称优于 Llama 3.2 3B、Qwen2.5 3B，无具体数字）。"
  },
  {
   "name": "IBM Granite",
   "layer": "m1",
   "url": "https://www.ibm.com/granite",
   "hq": "美国阿蒙克",
   "region": "海外",
   "type": "大厂产品线",
   "founded": 2023,
   "founded_month": "2023-09",
   "desc": "Apache 2.0 企业级小模型，3B 与 Nano 1B/350M 面向边缘",
   "modality": "语言",
   "products": [
    "Granite 4.1 3B",
    "Granite 4.0 Nano 1B",
    "Granite 4.0 Nano 350M"
   ],
   "params": {
    "v": "Granite 4.1 3B / 4.0 Nano 1B、350M（4.1 另有 8B、30B）",
    "src": "https://x.com/ArtificialAnlys/status/2049505499377193156"
   },
   "devices": [
    "iPhone",
    "Android",
    "Mac",
    "Windows",
    "Linux"
   ],
   "formats": [
    "GGUF",
    "LiteRT",
    "Core ML"
   ],
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://www.bigdata-insider.de/ibm-granite-4-1-open-source-llms-unternehmen-a-85af8135a975a9fee14e77cd583753ae/"
   },
   "benchmarks": [
    {
     "k": "Artificial Analysis Intelligence Index（Granite 4.1 3B）",
     "v": "9（8B 为 12，30B 为 15）",
     "date": "2026-04",
     "src": "https://x.com/ArtificialAnlys/status/2049505499377193156"
    }
   ],
   "key_data": [
    {
     "k": "iPhone 17 Pro 上 Granite 4.0-H 1B（int8）解码",
     "v": "约 30 tok/s（社区 CoreAI 转换页实测，非 IBM 官方）",
     "date": "2026-10",
     "src": "https://huggingface.co/mlboydaisuke/granite-4.0-h-CoreAI/blob/eef67e8fdd360d3fc99df1d05e358ed65b600619/README.md"
    }
   ],
   "product_line_of": "IBM",
   "public": true,
   "note": "Granite 4.1 于 2026-04-29 发布（Ertas 页）；4.0 Nano 1B/350M 于 2025-10 发布（AA 帖）。AA 称 Granite 4.1 8B 跑完指数只用 400 万输出 token，约为 Qwen3.5 9B 的 1/20。devices/formats 取自社区转换页与 Unsloth GGUF 发布，非 IBM 官方清单。"
  },
  {
   "name": "LG AI Research EXAONE 小尺寸",
   "layer": "m1",
   "url": "https://www.lgresearch.ai",
   "hq": "首尔",
   "region": "海外",
   "type": "研究机构",
   "founded": 2025,
   "founded_month": "2025-07",
   "desc": "EXAONE 4.0 的 1.2B 版，官方定位设备端应用",
   "modality": "语言",
   "products": [
    "EXAONE 4.0 1.2B"
   ],
   "params": {
    "v": "1.2B（另有 32B）",
    "src": "https://huggingface.co/OpenLLM-Korea/EXAONE-4.0-1.2B/blob/bc50b0169a8c6f7c914ee31b61684f5734f13860/README.md"
   },
   "license": {
    "v": "EXAONE AI Model License 1.2-NC（仅限研究与教育，非商用）",
    "kind": "开放权重",
    "src": "https://registry.ollama.com/sam860/exaone-4.0:1.2b-thinking"
   },
   "product_line_of": "LG",
   "note": "模型卡写明 1.2B 为 designed for on-device applications。设备、芯片、基准未查到。EXAONE 4.0 之后是否有新版小模型未查。"
  },
  {
   "name": "Meta Llama 小尺寸",
   "layer": "m1",
   "url": "https://www.llama.com",
   "hq": "美国门洛帕克",
   "region": "海外",
   "type": "大厂产品线",
   "founded": 2024,
   "founded_month": "2024-09",
   "desc": "Llama 3.2 1B/3B 剪枝蒸馏小模型，与高通、联发科优化，配 ExecuTorch",
   "modality": "语言",
   "products": [
    "Llama 3.2 1B",
    "Llama 3.2 3B"
   ],
   "params": {
    "v": "1B / 3B",
    "src": "https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/"
   },
   "devices": [
    "Android",
    "iPhone"
   ],
   "chips": [
    "Qualcomm Hexagon",
    "MediaTek APU"
   ],
   "formats": [
    "ExecuTorch",
    "GGUF"
   ],
   "license": {
    "v": "Llama 3.2 Community License（月活超 7 亿需另行授权）",
    "kind": "开放权重",
    "src": "https://huggingface.co/experimentalmachines/Llama-3.2-1B-Instruct-ExecuTorch/commit/62db024fe51112732f1934533f5cf8fa33c2221c"
   },
   "benchmarks": [
    {
     "k": "Llama 3.2 3B Instruct MMLU（5-shot）",
     "v": "63.4（1B 为 49.3）",
     "date": "2024-09",
     "src": "https://github.com/meta-llama/llama-models/blob/main/models/llama3_2/MODEL_CARD.md"
    }
   ],
   "product_line_of": "Meta",
   "public": true,
   "note": "Meta 称 3B 在指令遵循、摘要、工具调用上优于 Gemma 2 2.6B 和 Phi 3.5-mini，并与 Qualcomm、MediaTek 合作优化（Meta 博客）。是否有更新的 Llama 小尺寸版本未查到。设备列表来自初稿背景认知，未逐项核。"
  },
  {
   "name": "Microsoft Phi",
   "layer": "m1",
   "url": "https://azure.microsoft.com/en-us/products/phi",
   "hq": "美国雷德蒙德",
   "region": "海外",
   "type": "大厂产品线",
   "founded": 2023,
   "founded_month": "2023-06",
   "desc": "开放权重小模型系列 Phi，配 ONNX Runtime 在 PC 和手机本地跑",
   "modality": "语言",
   "products": [
    "Phi-4-mini-instruct 3.8B",
    "Phi-4-mini-reasoning 3.8B",
    "Phi-4-mini-flash-reasoning"
   ],
   "params": {
    "v": "3.8B（Phi-4-mini 系列）",
    "src": "https://aiwiki.ai/wiki/phi_4_mini"
   },
   "devices": [
    "Windows",
    "Android",
    "Linux",
    "Mac"
   ],
   "chips": [
    "Intel NPU",
    "Qualcomm Hexagon"
   ],
   "formats": [
    "ONNX",
    "GGUF"
   ],
   "license": {
    "v": "MIT",
    "kind": "开源",
    "src": "https://ai-solutions.wiki/tools/microsoft-phi/"
   },
   "benchmarks": [
    {
     "k": "Phi-4-mini-instruct MMLU（5-shot）",
     "v": "67.3",
     "date": "2025-02",
     "src": "https://aiwiki.ai/wiki/phi_4_mini"
    }
   ],
   "key_data": [
    {
     "k": "Q4_K_M GGUF 体积 / 最低内存",
     "v": "2.49 GB / 4 GB",
     "date": "2026-05",
     "src": "https://atomic.chat/models/phi-4-mini-instruct"
    }
   ],
   "product_line_of": "Microsoft",
   "public": true,
   "note": "Phi-4-mini 于 2025-02 发布。基准数字为 aiwiki 转录自 Hugging Face 模型卡，未直接打开模型卡核对。Phi-5 是否存在未查到。devices/chips 来自第一轮初稿背景认知（Copilot+ PC / ONNX Runtime GenAI），未逐项核。"
  },
  {
   "name": "Mistral AI（Ministral）",
   "layer": "m1",
   "url": "https://mistral.ai",
   "hq": "巴黎",
   "region": "海外",
   "type": "创业",
   "founded": 2023,
   "founded_month": "2023-04",
   "desc": "Ministral 3 的 3B/8B/14B 多模态小模型，3B 可在浏览器本地跑",
   "modality": "多模态",
   "products": [
    "Ministral 3 3B",
    "Ministral 3 8B",
    "Ministral 3 14B"
   ],
   "params": {
    "v": "3B / 8B / 14B",
    "src": "https://www.heise.de/en/news/Mistral-3-Four-new-AI-models-from-France-11101058.html"
   },
   "devices": [
    "浏览器",
    "Android"
   ],
   "formats": [
    "GGUF"
   ],
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://simonwillison.net/2025/Dec/2/introducing-mistral-3/"
   },
   "valuation": {
    "v": "超 €210 亿",
    "date": "2026-09",
    "kind": "投后估值，公司公布",
    "src": "https://www.euronews.com/2026/09/08/mistral-ai-raises-record-3-billion-in-samsung-led-funding-round"
   },
   "total_raised": "累计融资约 €58 亿（按上列五轮披露金额相加，非公司口径）",
   "funding": [
    {
     "date": "2023-06",
     "round": "种子轮",
     "amount": "€1.05 亿",
     "lead": "Lightspeed Venture Partners",
     "post": "约 €2.4 亿",
     "src": "https://www.dutchstartup.ai/en/news/what-is-mistral-ai-the-french-company-building-open-ai-models"
    },
    {
     "date": "2023-12",
     "round": "A 轮",
     "amount": "€3.85 亿",
     "lead": "Andreessen Horowitz",
     "post": "超 €20 亿",
     "src": "https://www.dutchstartup.ai/en/news/what-is-mistral-ai-the-french-company-building-open-ai-models"
    },
    {
     "date": "2024-06",
     "round": "B 轮",
     "amount": "€6 亿",
     "lead": "General Catalyst（跟投 Nvidia、IBM、Samsung、Bpifrance）",
     "src": "https://www.dutchstartup.ai/en/news/what-is-mistral-ai-the-french-company-building-open-ai-models"
    },
    {
     "date": "2025-09",
     "round": "C 轮",
     "amount": "€17 亿",
     "lead": "ASML（出资 €13 亿）",
     "post": "投后 €117 亿",
     "src": "https://www.orrick.com/zh-TW/News/2025/09/Orrick-conseille-Mistral-AI-dans-le-cadre-de-sa-levee-de-fonds-en-Serie-C"
    },
    {
     "date": "2026-09",
     "round": "D 轮",
     "amount": "€30 亿",
     "lead": "Samsung Electronics（跟投 Scaleup Europe Fund/EQT、PSG Equity）",
     "post": "超 €210 亿",
     "src": "https://www.euronews.com/2026/09/08/mistral-ai-raises-record-3-billion-in-samsung-led-funding-round"
    }
   ],
   "note": "端侧只是 Mistral 的副线（主线是云端大模型）。Ministral 3 于 2025-12-02 发布，14B 超出 10B，仅 3B/8B 属端侧范围。基准未查到：Artificial Analysis 小模型页有 Ministral 3 8B/3B 条目，数值未能确认。D 轮标题称 Series D，Mobile World Live 误写为 C 轮，以 Euronews/公司口径为准。设备仅 3B 浏览器 WebGPU 演示（Simon Willison 转述）。Voxtral 语音模型归 m3，不在此卡。"
  },
  {
   "name": "NVIDIA Nemotron Nano",
   "layer": "m1",
   "url": "https://developer.nvidia.com/nemotron",
   "hq": "美国圣克拉拉",
   "region": "海外",
   "type": "大厂产品线",
   "founded": 2025,
   "founded_month": "2025-08",
   "desc": "Mamba-Transformer 混合小模型，4B 版专为 Jetson 与本地 RTX 设备优化",
   "modality": "语言",
   "products": [
    "Nemotron 3 Nano 4B",
    "Nemotron Nano 9B v2"
   ],
   "params": {
    "v": "Nemotron 3 Nano 4B（约 3.97B，由 Nano 9B v2 压缩）/ Nano 9B v2",
    "src": "https://huggingface.co/blog/nvidia/nemotron-3-nano-4b"
   },
   "devices": [
    "Jetson",
    "Windows",
    "Linux"
   ],
   "chips": [
    "NVIDIA GPU"
   ],
   "formats": [
    "GGUF",
    "LiteRT"
   ],
   "license": {
    "v": "NVIDIA Open Model License（可商用、可做衍生模型）",
    "kind": "开放权重",
    "src": "https://huggingface.co/litert-community/NVIDIA-Nemotron-Nano-9B-v2/blame/main/LICENSE"
   },
   "key_data": [
    {
     "k": "Jetson Orin Nano 8GB 解码速度（Q4_K_M，llama.cpp）",
     "v": "18 tok/s（LinkedIn 转述 HF 模型卡）",
     "date": "2026-03",
     "src": "https://www.linkedin.com/posts/theaiconsultant_theaiconsultant-ai-edgeai-activity-7442526716071362560-jnY6"
    }
   ],
   "product_line_of": "NVIDIA",
   "public": true,
   "note": "Nemotron 3 Nano 4B 2026-03-17 发布，称是 NVIDIA 首个专为设备端部署优化的模型（HF 博客）。Nemotron 3 Nano 30B-A3B（总参数超 10B）不属本范围，未收。基准未查到可引用的一手数字（LinkedIn 转述的 MATH500 95.4% 等不采）。Nano 9B v2 有 LiteRT-LM 社区转换包。"
  },
  {
   "name": "TII Falcon-H1 小尺寸",
   "layer": "m1",
   "url": "https://falconllm.tii.ae",
   "hq": "阿布扎比",
   "region": "海外",
   "type": "研究机构",
   "founded": 2025,
   "desc": "Mamba+Transformer 混合架构小模型，含 90M Tiny 与 0.5B–3B",
   "modality": "语言",
   "products": [
    "Falcon-H1-Tiny 90M",
    "Falcon-H1 0.5B",
    "Falcon-H1 1.5B-Deep",
    "Falcon-H1 3B"
   ],
   "params": {
    "v": "Tiny 90M；Falcon-H1 0.5B / 1.5B / 3B",
    "src": "https://huggingface.co/litert-community/Falcon-H1-1.5B-Deep-Instruct/raw/d7636ad371a520f7bd84e16b9b4df7ca2ef7ee29/README.md"
   },
   "devices": [
    "Android",
    "Mac",
    "Linux"
   ],
   "formats": [
    "LiteRT",
    "GGUF",
    "MLX"
   ],
   "license": {
    "v": "Falcon LLM License",
    "kind": "开放权重",
    "src": "https://huggingface.co/tiiuae/Falcon-H1-Tiny-Coder-90M/blob/4c2bb4b084c8b4b078b6f2b6a72edd95e6431721/README.md"
   },
   "note": "Falcon-H1-Tiny 模型卡标 edge；LiteRT-LM 有 0.5B/1.5B/3B 转换包（litert-community）。Falcon-H1 首发时间、官方端侧设备清单未查到，founded 取 2025 为估计，需核。基准未查到。TII 为阿联酋政府研究机构，不披露融资。"
  },
  {
   "name": "南北阁 Nanbeige（BOSS直聘）",
   "layer": "m1",
   "url": "https://huggingface.co/Nanbeige",
   "hq": "北京",
   "region": "中国",
   "type": "大厂产品线",
   "founded": 2026,
   "founded_month": "2026-07",
   "desc": "BOSS直聘母公司 Kanzhun 的 3B 级开源小模型，主打小尺寸跑 Agent",
   "modality": "语言",
   "products": [
    "Nanbeige4.2-3B",
    "Nanbeige4.1-3B"
   ],
   "params": {
    "v": "Nanbeige4.2-3B：非嵌入参数 3B（总参数约 4B），Looped Transformer",
    "src": "https://hub-assets-cache.baai.ac.cn/view/56690"
   },
   "formats": [
    "GGUF",
    "MLX"
   ],
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://pandaily.com/nanbeige4-3-33b-open-source-agent-coding-model-512k-context-apache-2-boss-zhipin"
   },
   "benchmarks": [
    {
     "k": "Nanbeige4.2-3B SWE-Bench Verified / GPQA-Diamond（官方）",
     "v": "63.6 / 87.4",
     "date": "2026-07",
     "src": "https://hub-assets-cache.baai.ac.cn/view/56690"
    }
   ],
   "product_line_of": "Kanzhun（BOSS直聘）",
   "public": true,
   "note": "founded 取 4.2-3B 首发（2026-07）；Nanbeige4.1-3B 在 2026-02 已有社区转换版，团队早期发布时间未查。2026-10 发布 33B 教师模型 Nanbeige4.3-33B（非端侧，未列），称将蒸馏出 4.3-3B。设备、芯片未查到。此卡原在初稿 B 组（创业），因归属 Kanzhun 改作 m1 大厂产品线，请决定层级。"
  },
  {
   "name": "小米 MiMo",
   "layer": "m1",
   "url": "https://github.com/XiaomiMiMo",
   "hq": "北京",
   "region": "中国",
   "type": "大厂产品线",
   "founded": 2025,
   "founded_month": "2025-04",
   "desc": "MiMo-7B 推理小模型及面向终端的 MiMo-VL-Miloco，主线在 Agent 大模型",
   "modality": "多模态",
   "products": [
    "MiMo-7B",
    "MiMo-VL-Miloco"
   ],
   "params": {
    "v": "MiMo-7B 为 7B 稠密；MiMo-VL-Miloco 参数未查",
    "src": "https://ai.zol.com.cn/1101/11010043.html"
   },
   "license": {
    "v": "MiMo-7B：MIT",
    "kind": "开源",
    "src": "https://github.com/ariffazil/aaa/blob/fcc2d2f0abe2f3e400ccd63d5a5269a9eec66226/docs/research/MIMO_AI_DEEP_RESEARCH_2026-07-11.md"
   },
   "product_line_of": "小米",
   "public": true,
   "note": "MiMo-7B 2025-04-30 开源（百度百科），官方称在 AIME 24-25、LiveCodeBench v5 上超过 o1-mini（百科转述）。MiMo-V2-Flash 为 309B/15B 激活，不属端侧，未列入产品。MiMo-VL-Miloco 称面向终端设备（ZOL 报道），具体参数、设备未查到；小米手机上用的端侧模型与「澎湃 AI」的关系未查到。基准未查到一手来源。MIT 协议出自第三方整理文档（低权威）。端侧定位偏弱，如要精简可考虑不收。"
  },
  {
   "name": "智谱 GLM-Edge",
   "layer": "m1",
   "url": "https://www.zhipuai.cn",
   "hq": "北京",
   "region": "中国",
   "type": "创业",
   "founded": 2024,
   "founded_month": "2024-11",
   "desc": "GLM-Edge 端侧对话与视觉模型：1.5B/2B 面向手机车机，4B/5B 面向 PC",
   "modality": "多模态",
   "products": [
    "GLM-Edge-1.5B-Chat",
    "GLM-Edge-4B-Chat",
    "GLM-Edge-V-2B",
    "GLM-Edge-V-5B"
   ],
   "params": {
    "v": "1.5B / 4B（对话）、2B / 5B（视觉）",
    "src": "https://raw.githubusercontent.com/zai-org/GLM-Edge/refs/heads/main/README_en.md"
   },
   "devices": [
    "Android",
    "车机",
    "Windows"
   ],
   "chips": [
    "Qualcomm Hexagon"
   ],
   "formats": [
    "GGUF"
   ],
   "license": {
    "v": "GLM-4 许可：学术免费，商用需登记",
    "kind": "开放权重",
    "src": "https://huggingface.co/zai-org/glm-edge-1.5b-chat/commit/e55133ddaf63cf2c9a6b6647dbe13cb8e15e4169"
   },
   "valuation": {
    "v": "总市值约 2155 亿港元",
    "date": "2026-11",
    "kind": "二级市场市值，东方财富行情页",
    "src": "https://quote.eastmoney.com/hk/02513.html"
   },
   "total_raised": "IPO 前各轮私募未逐轮查；IPO 募资约 HK$43 亿（媒体口径）",
   "funding": [
    {
     "date": "2026-01",
     "round": "港股 IPO（02513.HK）",
     "amount": "约 HK$43 亿",
     "src": "https://www.chinaz.com/ainews/24392.shtml"
    }
   ],
   "key_data": [
    {
     "k": "骁龙 8 Elite 上解码速度（官方，混合量化）",
     "v": "1.5B 对话 / 2B 多模态 60+ tok/s，投机采样后峰值 100+ tok/s",
     "date": "2024-12",
     "src": "https://raw.githubusercontent.com/zai-org/GLM-Edge/refs/heads/main/README_en.md"
    }
   ],
   "public": true,
   "note": "智谱主线是云端 GLM 大模型，GLM-Edge 是 2024-11 发布的端侧系列；是否有 2025 年后的新端侧小尺寸未查到，如无则该卡偏旧。收录前请确认。IPO 的价格、募资额来自 CHINAZ，首日 2026-01-08 前后（报道发布日 2026-01-07）。私募轮次（含君联、联想创投等）本次未逐轮查。基准未查到。"
  },
  {
   "name": "百度 ERNIE 4.5 小模型",
   "layer": "m1",
   "url": "https://github.com/PaddlePaddle/ERNIE",
   "hq": "北京",
   "region": "中国",
   "type": "大厂产品线",
   "founded": 2025,
   "founded_month": "2025-06",
   "desc": "文心 4.5 开源系列中的 0.3B 稠密轻量模型，Apache 2.0 可商用",
   "modality": "语言",
   "products": [
    "ERNIE-4.5-0.3B"
   ],
   "params": {
    "v": "0.3B 稠密（同系列还有激活 3B 的 MoE，总参数更大，不属端侧）",
    "src": "https://www.peopleapp.com/column/30049553745-500006355719"
   },
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://www.peopleapp.com/column/30049553745-500006355719"
   },
   "product_line_of": "百度",
   "public": true,
   "note": "本家为补漏收录（初稿 A 组缺），理由：文心 4.5 于 2025-06-30 开源，含 0.3B 稠密模型，可算端侧入门尺寸。官方是否将其定位为端侧、设备/芯片支持、基准均未查到；如认为端侧定位不足可去掉。"
  },
  {
   "name": "腾讯混元小模型 Hunyuan",
   "layer": "m1",
   "url": "https://github.com/Tencent-Hunyuan",
   "hq": "深圳",
   "region": "中国",
   "type": "大厂产品线",
   "founded": 2025,
   "founded_month": "2025-08",
   "desc": "0.5B/1.8B/4B/7B 四款融合推理小模型，面向手机、座舱、智能家居",
   "modality": "语言",
   "products": [
    "Hunyuan-0.5B",
    "Hunyuan-1.8B",
    "Hunyuan-4B",
    "Hunyuan-7B"
   ],
   "params": {
    "v": "0.5B / 1.8B / 4B / 7B",
    "src": "https://stock.10jqka.com.cn/hks/20250805/c670138585.shtml"
   },
   "devices": [
    "Android",
    "Windows",
    "Mac",
    "车机"
   ],
   "chips": [
    "Qualcomm Hexagon",
    "MediaTek APU",
    "Intel NPU"
   ],
   "license": {
    "v": "开源，具体许可条款未查实",
    "kind": "开放权重",
    "src": "https://stock.10jqka.com.cn/hks/20250805/c670138585.shtml"
   },
   "product_line_of": "腾讯",
   "public": true,
   "note": "本家为补漏收录（初稿 A 组缺）。2025-08-04 开源；Arm、高通、Intel、联发科等芯片平台宣布支持部署（上证报转载）。官方称 0.5B 可在手机 CPU 上跑、7B 可在 RTX3060 笔记本跑（AITOP100 转述）。基准、具体许可未查到。"
  },
  {
   "name": "阿里 Qwen 小尺寸",
   "layer": "m1",
   "url": "https://qwen.ai",
   "hq": "杭州",
   "region": "中国",
   "type": "大厂产品线",
   "founded": 2023,
   "founded_month": "2023-08",
   "desc": "0.8B–9B 稠密多模态小模型，端侧最常用的开源底座之一",
   "modality": "多模态",
   "products": [
    "Qwen3.5-0.8B",
    "Qwen3.5-2B",
    "Qwen3.5-4B",
    "Qwen3.5-9B"
   ],
   "params": {
    "v": "0.8B / 2B / 4B / 9B",
    "src": "https://news.pconline.com.cn/2108/21082872.html"
   },
   "devices": [
    "Android",
    "iPhone",
    "树莓派",
    "Mac",
    "Linux"
   ],
   "formats": [
    "GGUF",
    "MLX"
   ],
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://news.pconline.com.cn/2108/21082872.html"
   },
   "benchmarks": [
    {
     "k": "Artificial Analysis Intelligence Index（推理模式）",
     "v": "9B 32 / 4B 27 / 2B 16 / 0.8B 9",
     "date": "2026-03",
     "src": "https://artificialanalysis.ai/pt/articles/qwen3-5-small-models"
    }
   ],
   "key_data": [
    {
     "k": "0.8B/2B 量化版体积",
     "v": "约 528 MB",
     "date": "2026-03",
     "src": "https://article.pchome.net/news/11343.html"
    }
   ],
   "product_line_of": "阿里巴巴",
   "note": "Qwen3.5 小尺寸 2026-03-02 开源；AA 称 9B 是 10B 以下最强、4B 是 5B 以下最强，但跑完指数需 2.3–3.9 亿输出 token。founded 取 Qwen 首次公开年份（2023）。设备列表来自 PChome（手机、IoT、树莓派）与初稿背景认知。"
  },
  {
   "name": "Arcee AI",
   "layer": "m2",
   "url": "https://www.arcee.ai",
   "hq": "旧金山",
   "region": "海外",
   "type": "创业",
   "founded": 2023,
   "founded_month": "2023-06",
   "desc": "从小模型定制起家，开放权重 Trinity 系列，Nano 6B 可端侧；主线已转 400B 前沿",
   "modality": "语言",
   "products": [
    "Trinity Nano 6B",
    "Trinity Mini 26B",
    "Trinity Large 400B"
   ],
   "params": {
    "v": "Trinity Nano 6B（Mini 26B、Large 400B 激活 13B 非端侧）",
    "src": "https://quantlogix.ai/private-company/arcee-ai"
   },
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://rushcommerce.dev/blog/arcee-ai-1b-valuation-open-weight-models-you-can-host"
   },
   "valuation": {
    "v": "超 $10 亿",
    "date": "2026-09",
    "kind": "B 轮估值，公司公布",
    "src": "https://siliconangle.com/2026/09/16/open-weight-model-developer-arcee-ai-reaches-1b-plus-valuation-with-undisclosed-series-b-funding/"
   },
   "total_raised": "累计融资未汇总披露（种子 + A 轮已披露 $2950 万，B 轮金额未披露）",
   "funding": [
    {
     "date": "2024-01",
     "round": "种子轮",
     "amount": "$550 万",
     "lead": "Long Journey Ventures（参与：Flybridge、Hugging Face CEO Clément Delangue）",
     "src": "https://www.arcee.ai/blog/small-language-models-rising-as-arcee-ai-lands-24m-series-a"
    },
    {
     "date": "2024-07",
     "round": "A 轮",
     "amount": "$2400 万",
     "lead": "Emergence Capital（跟投：Long Journey、Flybridge、Centre Street Partners、Arcadia Capital）",
     "src": "https://www.arcee.ai/blog/arcee-ai-secures-24m-series-a-to-transform-the-landscape-of-small-language-models"
    },
    {
     "date": "2026-09",
     "round": "B 轮",
     "amount": "未披露（Fortune 消息人士称至少 $1.5 亿）",
     "lead": "Vista Equity Partners、Cambium Capital、Emergence Capital（参与：M12、Hitachi、Wipro、IAG、AI10、P7）",
     "post": "投后超 $10 亿",
     "src": "https://siliconangle.com/2026/09/16/open-weight-model-developer-arcee-ai-reaches-1b-plus-valuation-with-undisclosed-series-b-funding/"
    }
   ],
   "note": "Trinity Nano（6B）与 Mini（26B）2025-12 发布。端侧设备、芯片、格式未查到官方说明；plainai.tech（低权威）称 Nano「small enough for on-device use」。基准未查到。主线已转 400B 前沿模型，端侧占比小。"
  },
  {
   "name": "Liquid AI",
   "layer": "m2",
   "url": "https://www.liquid.ai",
   "hq": "美国剑桥",
   "region": "海外",
   "type": "创业",
   "founded": 2023,
   "founded_month": "2023-03",
   "desc": "非 Transformer 的 LFM 混合架构，按目标硬件反推模型，做手机笔记本车载小模型",
   "modality": "多模态",
   "products": [
    "LFM2.5-2.6B",
    "LFM2.5-8B-A1B",
    "LFM2-24B-A2B",
    "LEAP / Apollo"
   ],
   "params": {
    "v": "LFM2.5-2.6B（约 2.7B）/ LFM2.5-8B-A1B（MoE，激活 1B）；系列从约 230M 到 40B+",
    "src": "https://capd.mit.edu/blog/2026/08/19/employer-spotlight-liquid-ai/"
   },
   "devices": [
    "Android",
    "iPhone",
    "Mac",
    "Windows",
    "车机"
   ],
   "chips": [
    "AMD XDNA"
   ],
   "formats": [
    "GGUF",
    "ONNX",
    "MLX"
   ],
   "license": {
    "v": "LFM Open License v1.0（自定义，非 OSI 开源）",
    "kind": "开放权重",
    "src": "https://www.orcarouter.ai/zh-CN/blog/kolibri-vs-lfm2-5-2-6b-base"
   },
   "benchmarks": [
    {
     "k": "LFM2.5-2.6B AIME25（公司自报）",
     "v": "51.87",
     "date": "2026-08",
     "src": "https://llm-stats.com/blog/research/lfm-2.5-2.6b-launch"
    }
   ],
   "valuation": {
    "v": "超 $20 亿",
    "date": "2024-12",
    "kind": "A 轮估值，公司员工在 MIT 页面自述；AlphaSignal 称 $23.5 亿",
    "src": "https://capd.mit.edu/blog/2026/08/19/employer-spotlight-liquid-ai/"
   },
   "total_raised": "累计融资约 $2.97 亿（PitchBook 口径）",
   "funding": [
    {
     "date": "2023-12",
     "round": "种子轮",
     "amount": "$4660 万（PitchBook）；另有口径 $3760 万",
     "lead": "OSS Capital、PagsGroup（参与：Automattic、Tom Preston-Werner 等）",
     "src": "https://pitchbook.com/profiles/company/541205-38"
    },
    {
     "date": "2024-12",
     "round": "A 轮",
     "amount": "$2.5 亿",
     "lead": "AMD Ventures（战略领投，跟投含 Shopify、OSS Capital、Duke Capital）",
     "post": "超 $20 亿（另有报道 $23.5 亿）",
     "src": "https://www.liquid.ai/blog/we-raised-250m-to-scale-capable-and-efficient-general-purpose-ai"
    }
   ],
   "key_data": [
    {
     "k": "LFM2.5-2.6B 手机级速度 / 内存",
     "v": "约 30 tok/s / 低于 2.5 GB（厂商自报）",
     "date": "2026-08",
     "src": "https://llm-stats.com/blog/research/lfm-2.5-2.6b-launch"
    }
   ],
   "customers": [
    "Mercedes-Benz",
    "Shopify",
    "Insilico Medicine"
   ],
   "note": "PitchBook 另列 2026-08 一笔金额未披露的 Early Stage VC，公司和主要媒体均未确认，未采。LFM2.5-2.6B 于 2026-08-04 发布。AMD XDNA 仅见与 AMD 合作优化的表述，未查到官方芯片支持清单。种子轮金额各源不一致（$37.6M / $37.5–46.6M）。"
  },
  {
   "name": "Moondream（M87 Labs）",
   "layer": "m2",
   "url": "https://moondream.ai",
   "hq": "西雅图",
   "region": "海外",
   "type": "创业",
   "founded": 2023,
   "founded_month": "2023-12",
   "desc": "小型视觉语言模型，能看图问答、检测、指点，原版小到可在手机上跑",
   "modality": "视觉",
   "products": [
    "Moondream 2（1.6B 级）",
    "Moondream 3 / 3.1（9B MoE，激活 2B）"
   ],
   "params": {
    "v": "Moondream 2：1.6B；Moondream 3.1：总 9B MoE、激活 2B",
    "src": "https://faun.dev/company/m87-labs/"
   },
   "devices": [
    "Android",
    "iPhone",
    "树莓派"
   ],
   "license": {
    "v": "moondream2 为 Apache-2.0；Moondream 3 系列为 source-available",
    "kind": "开放权重",
    "src": "https://faun.dev/toolbox/moondream/"
   },
   "total_raised": "累计融资约 $450 万（仅 Pre-seed）",
   "funding": [
    {
     "date": "2024-10",
     "round": "Pre-seed",
     "amount": "$450 万",
     "lead": "Felicis（参与：M12 GitHub Fund、Ascend）",
     "src": "https://www.geekwire.com/2024/seattle-startup-moondream-led-by-aws-vets-raises-4-5m-for-vision-language-model-software/"
    }
   ],
   "owner": "2026-10-13 被 Prime Intellect 收购（金额报道未见），两位联合创始人加入 Prime Intellect 负责推理工作",
   "note": "收购信息来自 GeekWire（引 The Information），员工少于 10 人。Moondream 2 的 1.6B 参数来自 GeekWire 2024 报道；Moondream 3.1（2026-07-07）参数与许可来自 FAUN（低权威）。树莓派、手机设备为 GeekWire/初稿表述，Moondream 3 系列的端侧运行能力未查。基准未查到。"
  },
  {
   "name": "Multiverse Computing",
   "layer": "m2",
   "url": "https://multiversecomputing.com",
   "hq": "西班牙圣塞巴斯蒂安",
   "region": "海外",
   "type": "创业",
   "founded": 2019,
   "desc": "CompactifAI 张量网络压缩，把大模型压到手机、PC、IoT 本地跑",
   "modality": "语言",
   "products": [
    "CompactifAI",
    "SuperFly（苍蝇脑）",
    "ChickBrain（鸡脑）"
   ],
   "params": {
    "v": "SuperFly / ChickBrain 参数未查到（公司称是为 IoT、手机、平板、PC 设计的超小模型）",
    "src": "https://ai.zhiding.cn/2025/0815/3170442.shtml"
   },
   "devices": [
    "iPhone",
    "Android",
    "Mac",
    "Windows"
   ],
   "license": {
    "v": "未查到统一协议",
    "src": "https://multiversecomputing.com"
   },
   "valuation": {
    "v": "$17 亿（投前）",
    "date": "2026-07",
    "kind": "C 轮投前估值，公司公布",
    "src": "https://multiversecomputing.com/resources/multiverse-computing-announces-series-c-fundraising-targeting-up-to-usd570m-eur500m-to-power"
   },
   "total_raised": "2025-08 创始人称自 2019 年累计约 $2.5 亿；公司预计 C 轮后总额约 $8 亿",
   "funding": [
    {
     "date": "2024",
     "round": "A 轮",
     "amount": "€2500 万",
     "lead": "Columbus Venture Partners（跟投：Quantonation、EIC Fund、Redstone QAI、Indi Partners）",
     "post": "约 €1 亿",
     "src": "https://columbusvp.com/es/columbus-vp-lidera-la-mayor-ronda-de-inversion-serie-a-en-software-cuantico-de-europa-al-financiar-a-multiverse-computing/"
    },
    {
     "date": "2025-06",
     "round": "B 轮",
     "amount": "€1.89 亿（$2.15 亿）",
     "lead": "Bullhound Capital（参与：HP Tech Ventures、SETT、Forgepoint、CDP VC、Toshiba 等）",
     "src": "https://multiversecomputing.com/resources/multiverse-computing-raises-usd215m-to-scale-ground-breaking-technology-that-compresses-llms-by"
    },
    {
     "date": "2026-07",
     "round": "C 轮（宣布目标，称 up to）",
     "amount": "$5.7 亿（€5 亿）",
     "lead": "Forgepoint Capital International、BNPP SIVF、Bullhound Capital 联合领投（参与：HP、Orange Ventures、Scania Invest、Santander 等）",
     "post": "投前 $17 亿（€15 亿）",
     "src": "https://multiversecomputing.com/resources/multiverse-computing-announces-series-c-fundraising-targeting-up-to-usd570m-eur500m-to-power"
    }
   ],
   "note": "A 轮年份只确认到 2024 年，月份未查。SuperFly/ChickBrain 于 2025-08 发布（至顶网转 TechCrunch）。公司主业是压缩技术与压缩后的模型库，不是从零训练的端侧模型公司，是否算 m2 请交回判断；模型许可、基准未查到。devices 仅取创始人称「可在 iPhone 甚至 Apple Watch 上本地运行」及发布稿的手机/平板/PC，未逐项核。HP 为投资方（AI PC 交叉）。"
  },
  {
   "name": "Pleias",
   "layer": "m2",
   "url": "https://pleias.ai",
   "hq": "巴黎",
   "region": "海外",
   "type": "创业",
   "founded": 2023,
   "founded_month": "2023-12",
   "desc": "法国小模型实验室，用许可清晰的合成数据训练极小推理模型",
   "modality": "语言",
   "products": [
    "Baguettotron 321M",
    "Monad 56M"
   ],
   "params": {
    "v": "Baguettotron 321M / Monad 56M",
    "src": "https://pleias.ai/blog/blogsynth-the-new-data-frontier"
   },
   "devices": [
    "Linux",
    "Windows",
    "Mac"
   ],
   "license": {
    "v": "未在检索结果中看到许可证条款",
    "src": "https://huggingface.co/PleIAs/Monad/blame/main/README.md"
   },
   "total_raised": "累计融资未披露（Tracxn、aiwiki 均标注未披露风险融资）",
   "note": "Tracxn 记 2024 年成立，法定公告显示 SAS 于 2023-12-26 设立，故 founded 取 2023。Monad 称是当前「最小可用」的 56M 小语言模型。设备仅有「可在消费级 GPU/CPU 本地推理」的 Tracxn 表述，具体手机、芯片未查到。基准未查到。"
  },
  {
   "name": "Sarvam AI",
   "layer": "m2",
   "url": "https://www.sarvam.ai",
   "hq": "班加罗尔",
   "region": "海外",
   "type": "创业",
   "founded": 2023,
   "founded_month": "2023-08",
   "desc": "印度语言全栈模型公司，Sarvam Edge 离线跑在手机、功能机、车和眼镜上",
   "modality": "多模态",
   "products": [
    "Sarvam Edge",
    "Sarvam-1（2B）"
   ],
   "params": {
    "v": "Sarvam-1：2B；Sarvam Edge 模型占用仅数 MB 级，参数量未披露",
    "src": "https://tech.yahoo.com/ai/meta-ai/articles/india-sarvam-wants-bring-ai-130104680.html"
   },
   "devices": [
    "Android",
    "车机"
   ],
   "license": {
    "v": "Sarvam-1：Sarvam AI Research License（arXiv 表格）；Sarvam Edge 未查到",
    "kind": "开放权重",
    "src": "http://export.arxiv.org/pdf/2501.13912"
   },
   "valuation": {
    "v": "$15 亿",
    "date": "2026-07",
    "kind": "投后估值，公司声明（经 VCCircle）",
    "src": "https://www.vccircle.com/sarvamai-joins-unicorn-club-with-234-mn-cheque-from-hcl-tech-others"
   },
   "total_raised": "累计融资约 $2.75 亿（A 轮 $4100 万 + B 轮首次交割 $2.34 亿，按 VCCircle 报道相加）",
   "funding": [
    {
     "date": "2023-12",
     "round": "A 轮",
     "amount": "$4100 万",
     "lead": "Lightspeed Venture Partners（参与：Peak XV、Khosla Ventures）",
     "src": "https://www.vccircle.com/sarvamai-joins-unicorn-club-with-234-mn-cheque-from-hcl-tech-others"
    },
    {
     "date": "2026-07",
     "round": "B 轮首次交割（计划总额 $3 亿）",
     "amount": "$2.34 亿（₹2,210 crore）",
     "lead": "HCLTech（参与：Bessemer、Khosla、Peak XV）",
     "post": "投后 $15 亿",
     "src": "https://www.vccircle.com/sarvamai-joins-unicorn-club-with-234-mn-cheque-from-hcl-tech-others"
    }
   ],
   "customers": [
    "HMD（Nokia 手机）"
   ],
   "note": "Sarvam Edge 2026-02-14 发布，展示离线语音识别、语音合成、翻译、文档数字化模型（CNBC-TV18）；2026-02 在 AI Impact Summit 宣布与 HMD 合作把对话助手放进 Nokia/HMD 手机，并计划上车和自家眼镜（Yahoo/TechCrunch 转载）。Edge 模型的参数、许可、基准均未公开。主线是云端全栈主权 AI，端侧是近期新增方向。"
  },
  {
   "name": "Zyphra",
   "layer": "m2",
   "url": "https://www.zyphra.com",
   "hq": "美国帕洛阿尔托",
   "region": "海外",
   "type": "创业",
   "desc": "Mamba2 混合架构 Zamba2 小模型（1.2B–7B）及 Zamba2-VL，近期转向前沿模型与云服务",
   "modality": "多模态",
   "products": [
    "Zamba2-1.2B",
    "Zamba2-2.7B",
    "Zamba2-VL",
    "Zonos（TTS）"
   ],
   "params": {
    "v": "Zamba2 1.2B / 2.7B / 7B；Zamba2-VL 含 2.7B",
    "src": "https://github.com/huggingface/transformers/blob/550d7b3834670483a4df436541272c055dc364bf/docs/source/en/model_doc/zamba2.md"
   },
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://github.com/huggingface/transformers/blob/550d7b3834670483a4df436541272c055dc364bf/docs/source/en/model_doc/zamba2.md"
   },
   "valuation": {
    "v": "$10 亿",
    "date": "2025",
    "kind": "A 轮估值，IBM 公告",
    "src": "https://www.techpowerup.com/341528/ibm-and-amd-collaborate-with-zyphra-on-next-generation-ai-infrastructure"
   },
   "total_raised": "累计约 $1.2 亿（ai-market-watch 口径，低权威）",
   "funding": [
    {
     "date": "2024-07",
     "round": "种子轮",
     "amount": "未披露",
     "src": "https://pitchbook.com/profiles/company/437883-40"
    },
    {
     "date": "2025-06",
     "round": "A 轮（时间两说：PitchBook 记 2025-06-09，IBM 2025-10-01 公告称「近期完成」）",
     "amount": "约 $1.0–1.1 亿（Forbes 引 PitchBook 为 $1.1 亿）",
     "lead": "未明确披露（参与：AMD、IBM、Future Ventures、Jaan Tallinn）",
     "post": "投后 $10 亿（IBM 公告）",
     "src": "https://www.techpowerup.com/341528/ibm-and-amd-collaborate-with-zyphra-on-next-generation-ai-infrastructure"
    }
   ],
   "key_data": [
    {
     "k": "Zamba2-VL 32k 预填首 token 延迟",
     "v": "比同类 Transformer 低约一个数量级（AI Weekly 转述）",
     "date": "2026-06",
     "src": "https://aiweekly.co/alerts/zyphra-zamba2-vl-cuts-time-to-first-token-10x"
    }
   ],
   "note": "Zamba2-7B 2024-07 发布，transformers 文档记 Zamba2 2024-11-22 发布。2026-05 Forbes/PitchBook 称其在洽 $5 亿 B 轮、估值 $50 亿以上（进行中，未确认，不入 funding）。公司已转向 Maia 超级助手与 Zyphra Cloud，端侧为早期方向。总部：多个来源写帕洛阿尔托，IBM 公告写旧金山；成立年份各源 2021/2023 不一，未写。设备、芯片未查到；AI Weekly 称其 Mamba2 内核需要 CUDA，纯 CPU 部署明显更慢。基准未查到可引用的单项。"
  },
  {
   "name": "元始智能 RWKV",
   "layer": "m2",
   "url": "https://www.rwkv.cn",
   "hq": "深圳",
   "region": "中国",
   "type": "创业",
   "founded": 2023,
   "founded_month": "2023-06",
   "desc": "RWKV 线性 RNN 架构开源模型，推理显存恒定，面向手机与 IoT 端侧",
   "modality": "语言",
   "products": [
    "RWKV-7 系列"
   ],
   "params": {
    "v": "RWKV-7 系列，具体端侧尺寸未查到",
    "src": "https://www.zgcsb.com/news/pinDao/2025-01/06/a_564099.html"
   },
   "license": {
    "v": "开源可商用（具体协议未查到）",
    "kind": "开源",
    "src": "https://www.zgcsb.com/news/pinDao/2025-01/06/a_564099.html"
   },
   "total_raised": "累计融资未汇总披露（数千万元天使轮 + 未披露种子轮）",
   "funding": [
    {
     "date": "2024-01",
     "round": "种子轮",
     "amount": "未披露",
     "lead": "奇绩创坛",
     "src": "https://www.chinaz.com/ainews/14451.shtml"
    },
    {
     "date": "2024-12",
     "round": "天使轮",
     "amount": "数千万元",
     "lead": "天际资本",
     "src": "https://www.chinaz.com/ainews/14451.shtml"
    }
   ],
   "note": "2026-01 另有「天使+轮，南山战新投独家投资」，但只见微信公众号文章，不采，请复核。资金用途含推进模型轻量化与端侧部署（移动端、IoT）。设备、芯片、基准未查到。"
  },
  {
   "name": "无界方舟 AutoArk",
   "layer": "m2",
   "url": "https://autoarkai.com",
   "hq": "上海",
   "region": "中国",
   "type": "创业",
   "desc": "Edge0 端侧 LLM、Audio8 语音模型与 EVA OS 硬件 AI 操作系统",
   "modality": "通用",
   "products": [
    "Edge0（端侧 LLM）",
    "Audio8（ASR / TTS）",
    "EVA OS"
   ],
   "params": {
    "v": "Edge0-8B-A1B-preview（8B MoE，端侧常驻内存约 1.0 GiB）/ Edge0-35B-A3B-preview（常驻约 2.9 GB）",
    "src": "https://github.com/Edge0-AI/Edge0/blob/main/README.md"
   },
   "devices": [
    "Mac",
    "iPhone"
   ],
   "chips": [
    "Apple ANE"
   ],
   "formats": [
    "MLX"
   ],
   "license": {
    "v": "Apache-2.0（Edge0-8B-A1B-preview 模型卡）",
    "kind": "开源",
    "src": "https://huggingface.co/Edge0/Edge0-8B-A1B-preview/blob/eff8b838a1042e1a97d1b48b4b7d864efe3c7732/README.md"
   },
   "benchmarks": [
    {
     "k": "edge0-8b（int4）OpenCompass 五项平均（自测）",
     "v": "69.9，对比 Ling 3.0 tiny fp16 的 72.7",
     "date": "2026-09",
     "src": "https://github.com/Edge0-AI/Edge0/blob/main/README.md"
    }
   ],
   "total_raised": "过去一年完成 4 轮融资，累计数亿元（易凯资本口径）",
   "funding": [
    {
     "date": "2026-04",
     "round": "Pre-A（连续两轮）",
     "amount": "未披露",
     "lead": "韶音、国瑞源基金、恒松资本、上海天使会",
     "src": "https://m.jrj.com.cn/madapter/24h/2026/04/02105956597568.shtml"
    }
   ],
   "key_data": [
    {
     "k": "edge0-8b 解码速度（Mac mini M4 Pro 24GB）",
     "v": "23.9–25.3 tok/s",
     "date": "2026-09",
     "src": "https://raw.githubusercontent.com/Edge0-AI/Edge0/master/README_zh.md"
    }
   ],
   "note": "仅按有一手出处的内容写：融资只写 2026-04 两轮 Pre-A；其后轮次与估值不写。Edge0 是流式 MoE 推理框架（SSD 专家卸载 + Recover-LoRA + 预路由），checkpoint 基于开源 MoE 底座（Qwen3.6-35B-A3B、Ling 3.0 tiny），README 已写明。Audio8 本次未单独查证。设备栏只取 README 实测的 Mac；手机运行为官方「phone-class 内存」表述，未见手机实测。成立年份未查。"
  },
  {
   "name": "面壁智能 ModelBest",
   "layer": "m2",
   "url": "https://modelbest.cn",
   "hq": "北京",
   "region": "中国",
   "type": "创业",
   "founded": 2022,
   "desc": "清华 NLP 系，MiniCPM「小钢炮」端侧开源模型，覆盖文本、视觉、全模态",
   "modality": "多模态",
   "products": [
    "MiniCPM 4.1-8B",
    "MiniCPM-V 4.5",
    "MiniCPM-o 4.5（全模态）"
   ],
   "params": {
    "v": "MiniCPM 系列 0.5B–9B",
    "src": "https://www.forbeschina.com/investment/%e6%8a%95%e8%b5%84/71360"
   },
   "devices": [
    "Android",
    "iPhone",
    "车机"
   ],
   "formats": [
    "GGUF"
   ],
   "license": {
    "v": "开源，具体协议各模型不同，未逐个核",
    "kind": "开源",
    "src": "https://www.forbeschina.com/investment/%e6%8a%95%e8%b5%84/71360"
   },
   "valuation": {
    "v": "超 200 亿元",
    "date": "2026-07",
    "kind": "媒体报道（智东西，经 C114 转载），公司未披露",
    "src": "https://www.c114.net.cn/industry/99944.html"
   },
   "total_raised": "2026 上半年累计融资超 50 亿元（媒体口径，智东西/凤凰网）；此前已披露 7 轮，2024-12 之前的轮次未逐一查",
   "funding": [
    {
     "date": "2024-12",
     "round": "数亿元（轮次名未披露）",
     "amount": "数亿元",
     "lead": "龙芯创投、鼎晖百孚、中关村科学城基金、赛富基金联合领投（跟投：北京市人工智能产业投资基金、清科创投）",
     "src": "https://finance.eastmoney.com/a/202602283657405361.html"
    },
    {
     "date": "2025-05",
     "round": "数亿元（轮次名未披露）",
     "amount": "数亿元",
     "lead": "洪泰基金、国中资本、清控金信、茅台基金联合投资",
     "src": "https://finance.eastmoney.com/a/202602283657405361.html"
    },
    {
     "date": "2025-12",
     "round": "数亿元（轮次名未披露）",
     "amount": "数亿元",
     "lead": "京国瑞、国科投资、中金保时捷基金、米聚资本、和基投资共同投出",
     "src": "https://finance.eastmoney.com/a/202602283657405361.html"
    },
    {
     "date": "2026-02",
     "round": "数亿元（轮次名未披露）",
     "amount": "数亿元",
     "lead": "中国电信（跟投：中信金石、中信私募）",
     "src": "https://stcn.com/article/detail/3655084.html"
    },
    {
     "date": "2026-04",
     "round": "数亿元（轮次名未披露）",
     "amount": "数亿元",
     "lead": "深创投、汇川产投联合领投（跟投：道禾长期投资、国泰君安创新投、武岳峰科创）",
     "src": "https://www.techweb.com.cn/finance/2026-04-07/2973953.shtml"
    },
    {
     "date": "2026-07",
     "round": "新一轮（轮次名、金额均未披露）",
     "amount": "未披露",
     "lead": "投资方为国家级基金、央企、汽车厂商等产业方及财务机构（具体名单见报道）",
     "src": "https://www.c114.net.cn/industry/99944.html"
    }
   ],
   "key_data": [
    {
     "k": "MiniCPM 累计下载量（GitHub + Hugging Face）",
     "v": "超 2400 万",
     "date": "2026-04",
     "src": "https://www.forbeschina.com/investment/%e6%8a%95%e8%b5%84/71360"
    }
   ],
   "note": "第一轮初稿写的「累计下载超 3800 万（至 2026-06-30）」和「2026-08 北京证监局辅导备案」本次没能重新打开出处，未采。与三星、长安马自达、吉利等的合作也未重新核，故 customers 留空。MiniCPM 具体开源协议未逐个查。基准未查到一手来源。"
  },
  {
   "name": "Alpha Cephei（Vosk）",
   "layer": "m3",
   "url": "https://alphacephei.com/vosk",
   "region": "海外",
   "type": "创业",
   "desc": "Vosk 离线语音识别工具包，50MB 小模型，树莓派和手机都能跑，覆盖 20+ 语言",
   "modality": "语音",
   "products": [
    "Vosk 小模型 / 大模型",
    "Vosk API（Python / Java / C# / JavaScript 绑定）"
   ],
   "devices": [
    "树莓派",
    "Android",
    "iPhone",
    "Linux",
    "Windows",
    "Mac"
   ],
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://github.com/alphacep/vosk-api"
   },
   "total_raised": "累计融资未披露",
   "note": "总部、成立时间、融资均未查到一手出处，已省略（初稿写俄罗斯 / 德国，未核）。基准未查到：不在 Open ASR 榜。语言数 20+、模型约 50MB 来自 PyPI 说明页 https://pypi.org/project/vosk/"
  },
  {
   "name": "Apple 设备端语音识别（SpeechAnalyzer）",
   "layer": "m3",
   "url": "https://developer.apple.com/documentation/speech/speechanalyzer",
   "hq": "库比蒂诺",
   "region": "海外",
   "type": "大厂产品线",
   "founded": 2025,
   "desc": "iOS 26 / macOS 26 内置的本地语音识别 API，音频不出设备，零调用成本",
   "modality": "语音",
   "products": [
    "SpeechAnalyzer",
    "SpeechTranscriber"
   ],
   "devices": [
    "iPhone",
    "Mac"
   ],
   "license": {
    "v": "系统内置，闭源",
    "kind": "闭源",
    "src": "https://developer.apple.com/documentation/speech/speechanalyzer"
   },
   "benchmarks": [
    {
     "k": "LibriSpeech test-clean / test-other WER（第三方评测，Lyonesse，非官方）",
     "v": "2.12% / 4.56%",
     "date": "2026-07",
     "src": "https://lyonesse.app/blog/apple-speech-api-benchmark.html"
    }
   ],
   "total_raised": "融资不适用（Apple 上市）",
   "product_line_of": "Apple",
   "public": true,
   "note": "补充收录（初稿漏）。Apple 官方未公布准确率；基准是第三方个人评测（5,559 句），仅作参考，不进 Open ASR 榜。"
  },
  {
   "name": "Applied Brain Research（ABR）",
   "layer": "m3",
   "url": "https://appliedbrainresearch.com",
   "hq": "滑铁卢",
   "region": "海外",
   "type": "创业",
   "desc": "状态空间模型 + 自研 TSP1 芯片，做几十 M 参数的端侧识别和合成，主打低功耗",
   "modality": "语音",
   "products": [
    "Niagara ASR（19M / 38M）",
    "Nith TTS",
    "TSP1 加速芯片",
    "ABR SDK"
   ],
   "params": {
    "v": "0.02B / 0.038B（Niagara ASR）",
    "src": "https://huggingface.co/abr-ai/niagara-38m-batch.en"
   },
   "devices": [
    "MCU"
   ],
   "chips": [
    "ABR TSP1"
   ],
   "license": {
    "v": "Applied Brain Research 开放许可",
    "kind": "开放权重",
    "src": "https://huggingface.co/abr-ai/niagara-38m-batch.en"
   },
   "benchmarks": [
    {
     "k": "Open ASR 平均 WER（Niagara-38M，第 58）",
     "v": "10.08%",
     "date": "2026-08",
     "src": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard"
    },
    {
     "k": "Open ASR 平均 WER（Niagara-19M，第 60）",
     "v": "12.50%",
     "date": "2026-08",
     "src": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard"
    }
   ],
   "total_raised": "累计融资未披露",
   "funding": [
    {
     "date": "2026-01",
     "round": "种子轮（超额认购）",
     "amount": "金额未披露",
     "lead": "Two Small Fish Ventures",
     "src": "https://www.semiconductor-digest.com/applied-brain-research-closes-seed-funding-round-to-accelerate-edge-ai-voice-processing/"
    }
   ],
   "key_data": [
    {
     "k": "ABR SDK / Niagara ASR / Nith TTS 通用可用",
     "v": "见 Distill 汇总",
     "date": "2026",
     "src": "https://www.distillintelligence.com/news/applied-brain-research"
    }
   ],
   "note": "Silicon Catalyst、Garage Capital、Raptor Group 等参与见 Distill 汇总页（低权威）。Nith TTS 5.5M 参数来自 Distill 汇总，未找到官方页。成立年份未核，已省略。官网对脚本返回 403，已用新闻稿 URL 交叉印证公司存在。"
  },
  {
   "name": "Argmax",
   "layer": "m3",
   "url": "https://www.argmaxinc.com",
   "hq": "旧金山",
   "region": "海外",
   "type": "创业",
   "founded": 2023,
   "founded_month": "2023-11",
   "desc": "把 Whisper、Parakeet 等优化到 Apple 芯片和骁龙上本地跑，WhisperKit 开源 + Pro SDK 按设备计费",
   "modality": "语音",
   "products": [
    "WhisperKit",
    "SpeakerKit",
    "TTSKit",
    "Argmax Pro SDK"
   ],
   "devices": [
    "iPhone",
    "Mac",
    "Android"
   ],
   "chips": [
    "Apple ANE",
    "Qualcomm Hexagon"
   ],
   "formats": [
    "Core ML",
    "LiteRT"
   ],
   "license": {
    "v": "MIT（WhisperKit 开源部分）；Pro SDK 商业",
    "kind": "开源",
    "src": "https://huggingface.co/argmaxinc/whisperkit-coreml"
   },
   "total_raised": "累计融资未披露",
   "funding": [
    {
     "date": "2024-11",
     "round": "未标明（Salesforce Ventures 官方页称其为参与方）",
     "amount": "金额未披露",
     "lead": "Salesforce Ventures、General Catalyst",
     "src": "https://salesforceventures.com/perspectives/welcome-argmax/"
    }
   ],
   "key_data": [
    {
     "k": "Hugging Face 近 30 天下载（argmaxinc/whisperkit-coreml）",
     "v": "31,110,151",
     "date": "2026-11",
     "src": "https://huggingface.co/argmaxinc/whisperkit-coreml"
    }
   ],
   "customers": [
    "Salesforce"
   ],
   "note": "基准未查到：不在 Open ASR 榜。成立时间取 FIRST CVC 页（2023-11），未见官方页；Simplify 页写累计 $8M，低权威不采。"
  },
  {
   "name": "Deepgram",
   "layer": "m3",
   "url": "https://deepgram.com",
   "hq": "旧金山",
   "region": "海外",
   "type": "创业",
   "founded": 2015,
   "desc": "云语音识别大厂，2026-07 起把 Nova-3 优化到骁龙 X 的 Hexagon NPU 上本地跑",
   "modality": "语音",
   "products": [
    "Nova-3（骁龙 PC 本地版）",
    "Flux",
    "Aura-2"
   ],
   "devices": [
    "Windows"
   ],
   "chips": [
    "Qualcomm Hexagon"
   ],
   "license": {
    "v": "商业闭源",
    "kind": "闭源",
    "src": "https://deepgram.com"
   },
   "valuation": {
    "v": "$13 亿",
    "date": "2026-01",
    "kind": "C 轮估值，公司公布",
    "src": "https://deepgram.com/learn/press-release-deepgram-raises-series-c"
   },
   "total_raised": "累计融资未取到一手汇总",
   "funding": [
    {
     "date": "2026-01",
     "round": "C 轮",
     "amount": "$1.3 亿",
     "lead": "AVP（Alkeon、In-Q-Tel、Madrona、Tiger、Wing、YC、BlackRock 管理基金等跟投）",
     "post": "估值 $13 亿",
     "src": "https://deepgram.com/learn/press-release-deepgram-raises-series-c"
    }
   ],
   "key_data": [
    {
     "k": "端侧发布：Nova-3 优化到 Snapdragon X 的 Hexagon NPU",
     "v": "2026-07-21 公司公告",
     "date": "2026-07",
     "src": "https://deepgram.com/learn/deepgram-delivers-real-time-voice-ai-at-the-edge-for-use-with-snapdragon"
    },
    {
     "k": "开发者数 / 付费组织",
     "v": "20 万+ 开发者；1,300+ 组织",
     "date": "2026-01",
     "src": "https://deepgram.com/learn/press-release-deepgram-raises-series-c"
    }
   ],
   "customers": [
    "Twilio"
   ],
   "note": "主线仍是云 API，端侧是 2026-07 起的新方向，收录依据是该公告。基准未查到：不在 Open ASR 榜；Nova-3 自报 WER 为自家口径，不采。"
  },
  {
   "name": "k2-fsa（新一代 Kaldi · sherpa-onnx）",
   "layer": "m3",
   "url": "https://github.com/k2-fsa/sherpa-onnx",
   "region": "中国",
   "type": "开源组织",
   "desc": "用 ONNX Runtime 在手机、嵌入式、NPU 上离线跑识别、合成、唤醒词的开源工具包",
   "modality": "语音",
   "products": [
    "sherpa-onnx v2.0.1（2026-10-26）",
    "新一代 Kaldi 预训练模型"
   ],
   "devices": [
    "Android",
    "iPhone",
    "树莓派",
    "Jetson",
    "Linux",
    "Windows",
    "Mac",
    "浏览器",
    "MCU"
   ],
   "chips": [
    "Qualcomm Hexagon",
    "Intel NPU",
    "Rockchip NPU",
    "Ascend NPU",
    "Axera NPU"
   ],
   "formats": [
    "ONNX"
   ],
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://github.com/k2-fsa/sherpa-onnx"
   },
   "total_raised": "累计融资未披露（非公司主体）",
   "customers": [
    "中国联通 Android App（关键词唤醒）",
    "中国人寿 App（关键词唤醒）",
    "MentraOS",
    "MediaTek Research BreezeApp"
   ],
   "note": "地区按文档与社区（微信 / QQ 群）推断，主要维护者所属机构未核，founded 省略。GitHub 星标因 API 限流未取到。基准未查到：本身是运行时，不是单一模型，不在 Open ASR 榜。客户为 README「Projects using」自述。同时覆盖 TTS / 唤醒词。"
  },
  {
   "name": "Moonshine AI（原 Useful Sensors）",
   "layer": "m3",
   "url": "https://moonshine.ai",
   "hq": "山景城",
   "region": "海外",
   "type": "创业",
   "founded": 2022,
   "desc": "为树莓派、手机做的低延迟流式端侧语音识别，开源 Moonshine Voice",
   "modality": "语音",
   "products": [
    "Moonshine Voice（2026-02）",
    "Moonshine Streaming tiny / small / medium",
    "Moonshine tiny / base"
   ],
   "params": {
    "v": "0.03B（streaming-tiny）/ 0.12B（streaming-small）/ 0.24B（streaming-medium）",
    "src": "https://huggingface.co/usefulsensors/moonshine-streaming-medium"
   },
   "devices": [
    "树莓派",
    "Android",
    "iPhone",
    "Mac",
    "Windows",
    "Linux"
   ],
   "license": {
    "v": "MIT",
    "kind": "开源",
    "src": "https://github.com/moonshine-ai/moonshine"
   },
   "benchmarks": [
    {
     "k": "Open ASR 平均 WER（streaming-medium，英文 65 模型中第 52）",
     "v": "7.69%",
     "date": "2026-08",
     "src": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard"
    },
    {
     "k": "Open ASR 平均 WER（streaming-small）",
     "v": "8.56%",
     "date": "2026-08",
     "src": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard"
    }
   ],
   "total_raised": "累计融资未披露",
   "funding": [
    {
     "date": "2025-07",
     "round": "种子轮",
     "amount": "金额未披露",
     "lead": "Wing VC、IQT（In-Q-Tel）",
     "src": "https://www.finsmes.com/2025/07/moonshine-ai-raises-funding.html"
    }
   ],
   "note": "融资金额公司未披露（CB Insights 页面写约 $5M，仅作线索，不采）。README 称 Python / iOS Swift Package 等绑定见 GitHub。Moonshine 论文 arXiv 2509.02523 讲多语小模型。"
  },
  {
   "name": "NVIDIA Parakeet / Canary 语音（NeMo · Riva）",
   "layer": "m3",
   "url": "https://developer.nvidia.com/nemo",
   "hq": "圣克拉拉",
   "region": "海外",
   "type": "大厂产品线",
   "desc": "开源高精度 ASR 系列，Riva 在 Jetson 上本地部署，社区已移植到手机",
   "modality": "语音",
   "products": [
    "Parakeet-TDT-0.6B v2 / v3",
    "Canary-Qwen-2.5B / Canary-1B-flash",
    "Nemotron Speech Streaming 0.6B",
    "Riva（Jetson）"
   ],
   "params": {
    "v": "0.6B（Parakeet）/ 1B（Canary-flash）/ 2.5B（Canary-Qwen）",
    "src": "https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3"
   },
   "devices": [
    "Jetson",
    "Linux",
    "Windows"
   ],
   "chips": [
    "NVIDIA GPU"
   ],
   "license": {
    "v": "CC-BY-4.0（Parakeet / Canary）",
    "kind": "开源",
    "src": "https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3"
   },
   "benchmarks": [
    {
     "k": "Open ASR 平均 WER（Canary-Qwen-2.5B，第 12）",
     "v": "5.63%",
     "date": "2026-08",
     "src": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard"
    },
    {
     "k": "Open ASR 平均 WER（Parakeet-TDT-0.6B v2，第 18）",
     "v": "5.88%",
     "date": "2026-08",
     "src": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard"
    },
    {
     "k": "Open ASR 平均 WER（Parakeet-TDT-0.6B v3，第 21）",
     "v": "6.04%",
     "date": "2026-08",
     "src": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard"
    }
   ],
   "key_data": [
    {
     "k": "Hugging Face 近 30 天下载（parakeet-tdt-0.6b-v3）",
     "v": "574,141",
     "date": "2026-11",
     "src": "https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3"
    },
    {
     "k": "Riva 2.26 嵌入式（L4T / JetPack 7.x）支持矩阵含 Parakeet、Nemotron ASR",
     "v": "见支持矩阵",
     "date": "2026-06",
     "src": "https://docs.nvidia.com/deeplearning/riva/user-guide/docs/support-matrix/support-matrix.html"
    }
   ],
   "product_line_of": "NVIDIA",
   "public": true,
   "note": "首次公开年份未核，founded 省略。融资不适用（NVIDIA 上市）。Nemotron-3.5-ASR-streaming-0.6B 在榜平均 WER 9.08%（第 56）。"
  },
  {
   "name": "OpenAI Whisper",
   "layer": "m3",
   "url": "https://github.com/openai/whisper",
   "hq": "旧金山",
   "region": "海外",
   "type": "大厂产品线",
   "founded": 2022,
   "founded_month": "2022-09",
   "desc": "开源 Whisper 是端侧语音识别的事实基线，被各移植版广泛本地运行",
   "modality": "语音",
   "products": [
    "Whisper large-v3",
    "Whisper large-v3-turbo",
    "Whisper tiny / base / small"
   ],
   "params": {
    "v": "large-v3-turbo 0.8B；large-v3 约 1.5B",
    "src": "https://huggingface.co/openai/whisper-large-v3-turbo"
   },
   "devices": [
    "iPhone",
    "Android",
    "Mac",
    "Windows",
    "Linux",
    "树莓派",
    "浏览器"
   ],
   "formats": [
    "Core ML",
    "ONNX"
   ],
   "license": {
    "v": "MIT（large-v3-turbo）/ Apache-2.0（large-v3）",
    "kind": "开源",
    "src": "https://huggingface.co/openai/whisper-large-v3-turbo"
   },
   "benchmarks": [
    {
     "k": "Open ASR 平均 WER（large-v3，英文 65 模型中第 43）",
     "v": "7.03%",
     "date": "2026-08",
     "src": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard"
    },
    {
     "k": "Open ASR 平均 WER（large-v3-turbo）",
     "v": "7.52%",
     "date": "2026-08",
     "src": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard"
    }
   ],
   "key_data": [
    {
     "k": "Hugging Face 近 30 天下载（large-v3-turbo）",
     "v": "5,636,560",
     "date": "2026-11",
     "src": "https://huggingface.co/openai/whisper-large-v3-turbo"
    },
    {
     "k": "Hugging Face 近 30 天下载（argmaxinc/whisperkit-coreml，Apple 端移植）",
     "v": "31,110,151",
     "date": "2026-11",
     "src": "https://huggingface.co/argmaxinc/whisperkit-coreml"
    }
   ],
   "product_line_of": "OpenAI",
   "note": "OpenAI 整体融资 / 估值不在此卡展开（非端侧专营）；本卡只记 Whisper 作为端侧基线。移植版 whisper.cpp 归运行时层（s2）。"
  },
  {
   "name": "Picovoice",
   "layer": "m3",
   "url": "https://picovoice.ai",
   "hq": "温哥华",
   "region": "海外",
   "type": "创业",
   "founded": 2018,
   "desc": "全离线语音栈：Cheetah 流式识别、Porcupine 唤醒词、Orca 合成，MCU 到浏览器都能跑",
   "modality": "语音",
   "products": [
    "Cheetah / Leopard（ASR）",
    "Porcupine（唤醒词）",
    "Orca（TTS）",
    "picoLLM"
   ],
   "devices": [
    "MCU",
    "树莓派",
    "Android",
    "iPhone",
    "Mac",
    "Windows",
    "Linux",
    "浏览器"
   ],
   "license": {
    "v": "闭源，需 AccessKey（含免费档）",
    "kind": "闭源",
    "src": "https://picovoice.ai"
   },
   "total_raised": "累计融资约 $50 万（仅见此一轮，之后未披露）",
   "funding": [
    {
     "date": "2020-12",
     "round": "天使 / Pre-seed",
     "amount": "$50 万",
     "lead": "Founders' Co-op",
     "src": "https://www.geekwire.com/tag/picovoice/"
    }
   ],
   "customers": [
    "Whirlpool",
    "LG"
   ],
   "note": "基准未查到：Picovoice 不在 Open ASR 榜上；其自有开源基准未核。员工 11–50 人（LinkedIn 公司页）。同时覆盖 m5 唤醒词，只放 m3 一张卡。"
  },
  {
   "name": "Speechmatics",
   "layer": "m3",
   "url": "https://www.speechmatics.com",
   "hq": "剑桥（英国）",
   "region": "海外",
   "type": "创业",
   "founded": 2006,
   "desc": "云端语音识别老牌，2026 起把近云端精度的模型搬到笔记本本地（Adobe Premiere）",
   "modality": "语音",
   "products": [
    "Adobe Premiere 本地语音转写模型（2026-04）",
    "Linden（2026-09）",
    "Enhanced / Melia 1"
   ],
   "devices": [
    "Windows",
    "Mac",
    "Linux"
   ],
   "license": {
    "v": "商业闭源",
    "kind": "闭源",
    "src": "https://www.speechmatics.com"
   },
   "benchmarks": [
    {
     "k": "Open ASR 平均 WER（speechmatics/enhanced，云 API，第 29）",
     "v": "6.29%",
     "date": "2026-08",
     "src": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard"
    },
    {
     "k": "Adobe 本地模型相对云端精度差距 / 1 小时音频耗时",
     "v": "5% 以内 / 约 55 秒",
     "date": "2026-04",
     "src": "https://www.linkedin.com/posts/speechmatics_adobe-x-speechmatice-on-device-activity-7452342882906185728-sHuL"
    }
   ],
   "total_raised": "累计融资约 $70M（Crustdata 聚合口径 $69.5M，B 轮公司口径 $62M）",
   "funding": [
    {
     "date": "2022-06",
     "round": "B 轮",
     "amount": "$62M",
     "lead": "Susquehanna Growth Equity（IQ Capital、AlbionVC 跟投）",
     "src": "https://www.speechmatics.com/company/articles-and-news/investment-funding-as-a-statement-of-belief"
    },
    {
     "date": "2019-10",
     "round": "A 轮",
     "amount": "$820 万",
     "lead": "IQ Capital（AlbionVC 跟投）",
     "src": "https://profiles.crustdata.com/company/speechmatics"
    }
   ],
   "key_data": [
    {
     "k": "Adobe 本地 STT 发布",
     "v": "2026-04-21 联合新闻稿",
     "date": "2026-04",
     "src": "https://www.globenewswire.com/de/news-release/2026/04/21/3278069/0/en/Adobe-and-Speechmatics-Deliver-Cloud-Grade-Speech-Recognition-On-Device-for-Premiere.html"
    },
    {
     "k": "端侧 STT 十问（官方文章）",
     "v": "2026-09-23",
     "date": "2026-09",
     "src": "https://www.speechmatics.com/company/articles-and-news/on-device-speech-to-text-in-10-questions"
    }
   ],
   "customers": [
    "Adobe"
   ],
   "note": "榜上条目是云端模型，端侧能力以 Adobe 合作和 2026-09 的本地部署发布为准。成立年份取聚合页「Est. 2006」，未找到一手出处。"
  },
  {
   "name": "科大讯飞 iFLYTEK 离线语音",
   "layer": "m3",
   "url": "https://www.xfyun.cn",
   "hq": "合肥",
   "region": "中国",
   "type": "大厂产品线",
   "desc": "讯飞开放平台的离线命令词识别与离线语音合成 SDK，嵌入式设备、机器人广泛使用",
   "modality": "语音",
   "products": [
    "离线命令词识别 SDK",
    "离线语音合成 SDK",
    "讯飞开放平台"
   ],
   "devices": [
    "Android",
    "iPhone",
    "MCU",
    "Linux"
   ],
   "license": {
    "v": "商业 SDK（按 appId 授权）",
    "kind": "闭源",
    "src": "https://www.xfyun.cn"
   },
   "revenue": {
    "v": "2025 年营收 271.05 亿元，同比 +16.12%（整个公司，非离线语音单项）",
    "date": "2025",
    "kind": "年报，公司公布",
    "src": "http://stock.10jqka.com.cn/20260429/c676374864.shtml"
   },
   "total_raised": "融资不适用（深交所上市 002230）",
   "key_data": [
    {
     "k": "第三方集成文档：iOS / Android 离线命令词识别 + 离线合成",
     "v": "见文档",
     "date": "2026-04",
     "src": "https://developer.yonyou.com/docs/Client-API/Art-Intel/IFlyVoice"
    }
   ],
   "product_line_of": "科大讯飞",
   "public": true,
   "note": "端侧能力来自 SDK 说明和第三方集成文档（用友开发者中心、Taobotics 机器人文档、讯飞官网），官网具体离线 SDK 页未逐页核。基准未查到；离线 SDK 为命令词级，不是大模型 ASR。首次公开年份未查到，founded 省略。"
  },
  {
   "name": "阿里通义语音 FunASR / SenseVoice / CosyVoice（含 Qwen3-ASR）",
   "layer": "m3",
   "url": "https://github.com/modelscope/FunASR",
   "hq": "杭州",
   "region": "中国",
   "type": "大厂产品线",
   "founded": 2024,
   "founded_month": "2024-07",
   "desc": "阿里开源中文语音识别与合成家族，Qwen3-ASR 最小 0.6B 面向端侧",
   "modality": "语音",
   "products": [
    "Qwen3-ASR 0.6B / 1.7B（2026-01）",
    "SenseVoice-Small",
    "FunASR 工具包",
    "Fun-CosyVoice3-0.5B"
   ],
   "params": {
    "v": "0.6B / 1.7B（Qwen3-ASR）；0.5B（CosyVoice3）",
    "src": "https://huggingface.co/Qwen/Qwen3-ASR-0.6B"
   },
   "devices": [
    "Android",
    "Linux",
    "Windows",
    "Mac"
   ],
   "license": {
    "v": "Apache-2.0（Qwen3-ASR、CosyVoice）；SenseVoice 用 FunASR 自有 Model License",
    "kind": "开源",
    "src": "https://huggingface.co/Qwen/Qwen3-ASR-0.6B"
   },
   "benchmarks": [
    {
     "k": "Open ASR 平均 WER（Qwen3-ASR-1.7B，第 7）",
     "v": "5.31%",
     "date": "2026-08",
     "src": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard"
    },
    {
     "k": "Open ASR 平均 WER（Qwen3-ASR-0.6B，第 23）",
     "v": "6.06%",
     "date": "2026-08",
     "src": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard"
    }
   ],
   "total_raised": "融资不适用（阿里巴巴）",
   "key_data": [
    {
     "k": "Hugging Face 近 30 天下载（Qwen3-ASR-1.7B）",
     "v": "1,614,839",
     "date": "2026-11",
     "src": "https://huggingface.co/Qwen/Qwen3-ASR-1.7B"
    },
    {
     "k": "Hugging Face 近 30 天下载（Fun-CosyVoice3-0.5B-2512）",
     "v": "260,255",
     "date": "2026-11",
     "src": "https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512"
    },
    {
     "k": "Qwen3-ASR 开源（52 种语言与方言，最小 0.6B）",
     "v": "2026-01-29",
     "date": "2026-01",
     "src": "https://www.qbitai.com/2026/01/374421.html"
    }
   ],
   "product_line_of": "阿里巴巴",
   "public": true,
   "note": "基准里的 Qwen3-ASR 来自阿里 Qwen 团队，与 FunASR / SenseVoice 是不同团队的模型，合并在一张阿里语音卡里；founded 取 SenseVoice / CosyVoice 开源年（2024-07 取自 sherpa-onnx 模型命名，未找到官方发布页，弱）。SenseVoice-Small 参数量未取到。Qwen-Omni 另有一张 m5 卡。"
  },
  {
   "name": "Canopy Labs（Orpheus）",
   "layer": "m4",
   "url": "https://canopylabs.ai",
   "region": "海外",
   "type": "创业",
   "desc": "开源 Orpheus TTS（Llama 底座，3B 起另有小尺寸），可本地量化部署",
   "modality": "语音",
   "products": [
    "Orpheus TTS 3B"
   ],
   "params": {
    "v": "3B（另有更小尺寸）",
    "src": "https://huggingface.co/canopylabs/orpheus-3b-0.1-ft"
   },
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://huggingface.co/canopylabs/orpheus-3b-0.1-ft"
   },
   "total_raised": "累计融资未披露",
   "key_data": [
    {
     "k": "Hugging Face 近 30 天下载（orpheus-3b-0.1-ft）",
     "v": "32,003",
     "date": "2026-11",
     "src": "https://huggingface.co/canopylabs/orpheus-3b-0.1-ft"
    }
   ],
   "note": "拿不准端侧占比：官方定位偏数字人 / 头像，Orpheus 是 3B 级开放权重，手机部署证据没查到。总部、成立时间、融资未查到一手（arts.talk 写 $17M，来源不明，不采）。小尺寸「150M 到 3B」来自 Second Talent 文章（低权威）。"
  },
  {
   "name": "Cartesia",
   "layer": "m4",
   "url": "https://www.cartesia.ai",
   "hq": "旧金山",
   "region": "海外",
   "type": "创业",
   "founded": 2023,
   "desc": "状态空间模型做实时语音，Sonic 云 API 之外有 Sonic On-Device 与开源 Edge 库",
   "modality": "语音",
   "products": [
    "Sonic-3 / Sonic-3.5",
    "Ink（STT）",
    "Sonic On-Device（私测）",
    "Edge 库 / Rene 1.3B"
   ],
   "params": {
    "v": "Rene 1.3B（开源端侧 SSM 语言模型）",
    "src": "https://www.cartesia.ai/blog/on-device"
   },
   "devices": [
    "Mac",
    "iPhone"
   ],
   "chips": [
    "Apple ANE"
   ],
   "formats": [
    "MLX"
   ],
   "license": {
    "v": "Sonic 闭源；Edge 库 / Rene Apache-2.0",
    "kind": "闭源",
    "src": "https://www.cartesia.ai/blog/on-device"
   },
   "total_raised": "累计融资约 $1.64 亿（A+B；聚合页另有 $1.91 亿含种子，口径不一）",
   "funding": [
    {
     "date": "2025-10",
     "round": "B 轮",
     "amount": "$1 亿",
     "lead": "Kleiner Perkins（Index Ventures、Lightspeed、NVIDIA 跟投）",
     "src": "https://www.hindustantimes.com/trending/us/who-is-karan-goel-dps-rk-puram-and-iit-delhi-graduate-who-raised-100-million-for-his-silicon-valley-startup-101761731790909.html"
    },
    {
     "date": "2025-03",
     "round": "A 轮",
     "amount": "$6,400 万",
     "lead": "Kleiner Perkins",
     "src": "https://neolabs.gtmascode.dev/labs/cartesia/"
    }
   ],
   "key_data": [
    {
     "k": "端侧首发：Edge 库 + Rene + Sonic On-Device（私测）",
     "v": "2024-08-28",
     "date": "2024-08",
     "src": "https://www.cartesia.ai/blog/on-device"
    }
   ],
   "customers": [
    "Tencent Cloud（战略合作）"
   ],
   "note": "估值未披露（聚合页 $1.2B 无出处，不采）；A 轮来源为聚合页（低权威），B 轮来源为 CEO 在 X 上的公告被 Hindustan Times 转述。Sonic On-Device 目标硬件为 Apple 芯片；是否已 GA 未查到。基准未查到官方可比榜。"
  },
  {
   "name": "Coqui",
   "layer": "m4",
   "url": "https://github.com/coqui-ai/TTS",
   "hq": "柏林",
   "region": "海外",
   "type": "创业",
   "founded": 2021,
   "desc": "开源 TTS（XTTS）出身的创业公司，2024-01 宣布关闭，代码被社区 fork 延续",
   "modality": "语音",
   "products": [
    "XTTS-v2",
    "Coqui TTS"
   ],
   "params": {
    "v": "467M（XTTS-v2）",
    "src": "https://huggingface.co/hexgrad/Kokoro-82M"
   },
   "license": {
    "v": "CPML（XTTS-v2，非商用）；代码 MPL-2.0",
    "kind": "开放权重",
    "src": "https://github.com/idiap/coqui-ai-TTS/discussions/216"
   },
   "total_raised": "累计融资约 $330 万（澎湃新闻转述）",
   "funding": [
    {
     "date": "2023-03",
     "round": "种子轮",
     "amount": "$330 万",
     "lead": "未取到",
     "src": "https://m.thepaper.cn/newsDetail_forward_26400053"
    }
   ],
   "key_data": [
    {
     "k": "关闭声明（创始人 LinkedIn 帖）",
     "v": "2024-01-02",
     "date": "2024-01",
     "src": "https://www.linkedin.com/posts/erengolge_coqui-is-shutting-down-thanks-for-all-activity-7148322600681701376-w2ch"
    }
   ],
   "owner": "2024-01 宣布关闭，代码由 idiap/coqui-ai-TTS 等社区 fork 维护",
   "note": "已关闭，仅留作图谱里的历史节点；也可由你决定是否删除。XTTS-v2 的 467M 取自 Kokoro 模型卡的对比表；柏林总部取自背景认知，澎湃稿只写出身 Mozilla，请当 ○ 对待。"
  },
  {
   "name": "Fish Audio",
   "layer": "m4",
   "url": "https://fish.audio",
   "hq": "帕洛阿尔托",
   "region": "海外",
   "type": "创业",
   "founded": 2025,
   "desc": "由开源 Fish Speech 起家的语音生成公司，开放权重可自托管，主线是云 API",
   "modality": "语音",
   "products": [
    "Fish Audio S2-Pro",
    "OpenAudio S1-mini（0.5B）",
    "Fish Speech（开源）"
   ],
   "license": {
    "v": "Fish Audio Research License（S2-Pro，非商用研究）",
    "kind": "开放权重",
    "src": "https://huggingface.co/fishaudio/s2-pro"
   },
   "revenue": {
    "v": "ARR 约 $2,100 万",
    "date": "2026-07",
    "kind": "公司口径（融资新闻稿）",
    "src": "https://www.tipranks.com/news/private-companies/fish-audio-secures-52-million-seed-round-as-arr-hits-21-million-in-first-year"
   },
   "total_raised": "累计融资约 $5,200 万（仅种子轮）",
   "funding": [
    {
     "date": "2026-07",
     "round": "种子轮",
     "amount": "$5,200 万",
     "lead": "Coreline Ventures、Capital Today（359 Capital、Play Time、HF0、645 Ventures、Parable 等参与）",
     "src": "https://digital-release.kxan.com/business/press-releases/cision/20260728LA13051/fish-audio-raises-52m-in-seed-funding-after-turning-passion-project-into-one-of-voice-ais-fastest-growing-companies/"
    }
   ],
   "key_data": [
    {
     "k": "用户数（开源 + 托管）",
     "v": "800 万+",
     "date": "2026-07",
     "src": "https://finance.yahoo.com/technology/ai/articles/fish-audio-raises-50m-seed-140000817.html"
    },
    {
     "k": "Hugging Face 近 30 天下载（s2-pro）",
     "v": "48,306",
     "date": "2026-11",
     "src": "https://huggingface.co/fishaudio/s2-pro"
    }
   ],
   "note": "拿不准是否算端侧：权重可自托管，但目标硬件是 GPU，未查到手机 / MCU 部署；S2-Pro 许可为研究许可。OpenAudio S1-mini 0.5B 来自 audio-edge-index 记录，本轮未打开官方模型卡（HF 路径 404）。建议交回由你决定是否保留。"
  },
  {
   "name": "Gradium",
   "layer": "m4",
   "url": "https://gradium.ai",
   "hq": "巴黎",
   "region": "海外",
   "type": "创业",
   "founded": 2025,
   "founded_month": "2025-09",
   "desc": "Kyutai 分拆公司，云端语音 API 之外于 2026-10 开源 31M / 40M 的端侧 TTS Phonon",
   "modality": "语音",
   "products": [
    "Phonon 31M / 40M（端侧 TTS，2026-10）",
    "TTS / STT / 语音克隆 API"
   ],
   "params": {
    "v": "0.031B / 0.04B",
    "src": "https://huggingface.co/gradium/phonon-31m-en"
   },
   "devices": [
    "Mac",
    "Windows",
    "Linux"
   ],
   "license": {
    "v": "Apache-2.0（Phonon 31M 英语）；40M 系列 CC-BY-NC-4.0",
    "kind": "开放权重",
    "src": "https://huggingface.co/gradium/phonon-31m-en"
   },
   "total_raised": "累计融资超 €1 亿（公司口径，2026-07）",
   "funding": [
    {
     "date": "2025-12",
     "round": "种子轮",
     "amount": "€6,000 万",
     "lead": "FirstMark Capital、Eurazeo（Niel、Saadé、Schmidt、Korelya Capital、Amplify Partners、DST 等跟投）",
     "src": "https://www.channelnews.fr/gradium-spin-off-de-kyutai-leve-60-millions-deuros-pour-conquerir-lia-vocale-153234"
    },
    {
     "date": "2026-07",
     "round": "种子轮延伸",
     "amount": "超 €4,000 万（总额超 €1 亿）",
     "lead": "NVIDIA 新加入",
     "src": "https://www.latribune.fr/article/tech/17414600668399/gradium-passe-la-barre-des-100-millions-deuros-avec-le-soutien-de-nvidia"
    }
   ],
   "key_data": [
    {
     "k": "Hugging Face 近 30 天下载（phonon-31m-en）",
     "v": "120,463",
     "date": "2026-11",
     "src": "https://huggingface.co/gradium/phonon-31m-en"
    }
   ],
   "note": "主线是云 API，端侧是 2026-10 的 Phonon 开放权重（Phonon 的发布细节来自 HF 页，官方博客未单独打开）。基准未查到。"
  },
  {
   "name": "KittenML（Kitten TTS）",
   "layer": "m4",
   "url": "https://github.com/KittenML/KittenTTS",
   "region": "海外",
   "type": "创业",
   "founded": 2025,
   "founded_month": "2025-08",
   "desc": "15M 起步的超小开源 TTS，25–80MB，无需 GPU，树莓派级设备可跑",
   "modality": "语音",
   "products": [
    "Kitten TTS nano 15M",
    "Kitten TTS micro 40M",
    "Kitten TTS mini 80M"
   ],
   "params": {
    "v": "15M / 40M / 80M",
    "src": "https://huggingface.co/KittenML/kitten-tts-nano-0.8-fp32"
   },
   "devices": [
    "树莓派",
    "Linux",
    "Windows",
    "Mac"
   ],
   "formats": [
    "ONNX",
    "LiteRT"
   ],
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://huggingface.co/KittenML/kitten-tts-nano-0.8-fp32"
   },
   "total_raised": "累计融资未披露",
   "key_data": [
    {
     "k": "Hugging Face 近 30 天下载（kitten-tts-nano-0.8-fp32）",
     "v": "253,307",
     "date": "2026-11",
     "src": "https://huggingface.co/KittenML/kitten-tts-nano-0.8-fp32"
    },
    {
     "k": "Hugging Face 近 30 天下载（kitten-tts-mini-0.8）",
     "v": "66,820",
     "date": "2026-11",
     "src": "https://huggingface.co/KittenML/kitten-tts-mini-0.8"
    }
   ],
   "note": "补充收录（初稿漏）。公司主体、总部未查，省略；首发时间取 Korben 报道（2025-08）。LiteRT 版是社区转换（litert-community）。espeak-ng 前端为 GPL-3.0，见 LiteRT 转换页说明。基准未查到。"
  },
  {
   "name": "Kokoro（hexgrad）",
   "layer": "m4",
   "url": "https://huggingface.co/hexgrad/Kokoro-82M",
   "region": "海外",
   "type": "开源组织",
   "founded": 2024,
   "founded_month": "2024-12",
   "desc": "82M 参数的开源 TTS，体积小、CPU 和手机上都能离线实时合成",
   "modality": "语音",
   "products": [
    "Kokoro-82M v1.0"
   ],
   "params": {
    "v": "0.082B",
    "src": "https://huggingface.co/hexgrad/Kokoro-82M"
   },
   "devices": [
    "Android",
    "Mac",
    "Windows",
    "Linux"
   ],
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://huggingface.co/hexgrad/Kokoro-82M"
   },
   "benchmarks": [
    {
     "k": "Artificial Analysis 文本转语音榜 Elo（开放权重榜第 3）",
     "v": "1057",
     "date": "2026-11",
     "src": "https://artificialanalysis.ai/text-to-speech/leaderboard?tab=leaderboard"
    },
    {
     "k": "v0.19 发布时在 TTS Spaces Arena 单声线榜排第 1",
     "v": "第 1",
     "date": "2024-12",
     "src": "https://huggingface.co/hexgrad/Kokoro-82M"
    }
   ],
   "total_raised": "无公司主体，融资不适用",
   "key_data": [
    {
     "k": "Hugging Face 近 30 天下载",
     "v": "11,535,189",
     "date": "2026-11",
     "src": "https://huggingface.co/hexgrad/Kokoro-82M"
    },
    {
     "k": "Android 离线 TTS App 使用 Kokoro-82M（sherpa-onnx 项目列表）",
     "v": "VoxSherpa TTS",
     "date": "2026-11",
     "src": "https://github.com/k2-fsa/sherpa-onnx"
    }
   ],
   "note": "主体未查到（个人项目），hq、region 省略。founded 取 v0.19 权重发布（2024-12-25，模型卡）。"
  },
  {
   "name": "Neuphonic",
   "layer": "m4",
   "url": "https://neuphonic.com",
   "hq": "伦敦",
   "region": "海外",
   "type": "创业",
   "desc": "开源 NeuTTS：基于 0.5B 级语言模型的 TTS，GGUF 量化后在普通 CPU 上实时合成并克隆声线",
   "modality": "语音",
   "products": [
    "NeuTTS Air",
    "NeuTTS Nano",
    "NeuCodec"
   ],
   "params": {
    "v": "约 0.75B（NeuTTS Air，HF 统计）",
    "src": "https://huggingface.co/neuphonic/neutts-air"
   },
   "devices": [
    "Windows",
    "Mac",
    "Linux"
   ],
   "formats": [
    "GGUF"
   ],
   "license": {
    "v": "Apache-2.0（NeuTTS Air）",
    "kind": "开源",
    "src": "https://huggingface.co/neuphonic/neutts-air"
   },
   "total_raised": "累计融资约 £300 万（Tracxn 记 $3.91M，同一轮）",
   "funding": [
    {
     "date": "2024-10",
     "round": "Pre-seed",
     "amount": "£300 万",
     "lead": "Moonfire VC（Tiny VC、Salica Oryx Fund、Cur8 Capital 跟投）",
     "src": "https://www.finsmes.com/2024/10/neuphonic-raises-3m-in-pre-seed-funding.html"
    }
   ],
   "key_data": [
    {
     "k": "Hugging Face 点赞数（neutts-air）",
     "v": "909",
     "date": "2026-11",
     "src": "https://huggingface.co/neuphonic/neutts-air"
    }
   ],
   "note": "成立年份冲突（PitchBook 2021，Tracxn 2024），已省略。融资公告称资金用于「开发端侧方案」。GGUF 与 Qwen 0.5B 底座说法来自 exploreai.directory（低权威）：https://www.exploreai.directory/tools/neuphonic ；参数量取 HF 统计，与「0.5B」不一致，以 HF 为准。基准未查到。"
  },
  {
   "name": "Piper（Open Home Foundation）",
   "layer": "m4",
   "url": "https://github.com/OHF-Voice/piper1-gpl",
   "region": "海外",
   "type": "开源组织",
   "founded": 2025,
   "desc": "Home Assistant 默认的离线 TTS，树莓派上 CPU 实时合成",
   "modality": "语音",
   "products": [
    "Piper（piper1-gpl，v1.8.0 于 2026-09-04）",
    "piper-voices 语音包"
   ],
   "devices": [
    "树莓派",
    "Linux",
    "Windows",
    "Mac"
   ],
   "license": {
    "v": "GPL-3.0（piper1-gpl；原 rhasspy/piper 为 MIT，已于 2025-10 归档）",
    "kind": "开源",
    "src": "https://github.com/OHF-Voice/piper1-gpl"
   },
   "total_raised": "非营利基金会，融资不适用",
   "key_data": [
    {
     "k": "piper-voices Hugging Face 点赞数",
     "v": "846",
     "date": "2026-11",
     "src": "https://huggingface.co/rhasspy/piper-voices"
    }
   ],
   "note": "founded 取现维护仓库创建时间（2025-03-28，据评测文章转述）；Piper 本体由 Rhasspy 项目更早开发，具体首发年未查。hq 具体城市未查。基准未查到：无公开榜单。"
  },
  {
   "name": "Resemble AI（Chatterbox）",
   "layer": "m4",
   "url": "https://www.resemble.ai",
   "hq": "旧金山",
   "region": "海外",
   "type": "创业",
   "founded": 2019,
   "desc": "开源 Chatterbox（0.5B）可在本地跑的声线克隆 TTS；公司主营云端语音与深度伪造检测",
   "modality": "语音",
   "products": [
    "Chatterbox",
    "Chatterbox Multilingual",
    "Detect"
   ],
   "params": {
    "v": "0.5B（Chatterbox）",
    "src": "https://modal.com/blog/open-source-tts"
   },
   "license": {
    "v": "MIT",
    "kind": "开源",
    "src": "https://huggingface.co/ResembleAI/chatterbox"
   },
   "total_raised": "累计融资约 $2,500 万（SiliconANGLE）",
   "funding": [
    {
     "date": "2025-12",
     "round": "轮次各源不一（CB Insights 记 B 轮，Seedtable 记 A 轮）",
     "amount": "$1,300 万",
     "lead": "未点名（Okta Ventures、Sony Group Innovation Fund、Google、Comcast Ventures、Craft Ventures 等参与）",
     "src": "https://siliconangle.com/2025/12/08/resemble-ai-hauls-13m-different-approach-deepfake-detection/"
    }
   ],
   "key_data": [
    {
     "k": "Hugging Face 近 30 天下载（chatterbox）",
     "v": "2,239,938",
     "date": "2026-11",
     "src": "https://huggingface.co/ResembleAI/chatterbox"
    }
   ],
   "note": "拿不准端侧占比：公司主线是云端和检测，Chatterbox 是开源模型。SiliconANGLE 页未写轮次名。成立年份取 Seedtable 首轮 2019，hq 未找到一手出处。2026-03 另有 Wa'ed Ventures 战略轮（Caplight，低权威）未采。"
  },
  {
   "name": "Sesame",
   "layer": "m4",
   "url": "https://www.sesame.com",
   "hq": "旧金山",
   "region": "海外",
   "type": "创业",
   "founded": 2023,
   "desc": "拟人对话语音模型 CSM（1B 开源）加自研 AI 眼镜；对话产品 Maya / Miles 在云端",
   "modality": "语音",
   "products": [
    "CSM-1B（开源）",
    "Maya / Miles 语音伙伴",
    "AI 眼镜（在研）"
   ],
   "params": {
    "v": "1B（CSM-1B）",
    "src": "https://huggingface.co/sesame/csm-1b"
   },
   "devices": [
    "iPhone"
   ],
   "license": {
    "v": "Apache-2.0（CSM-1B）",
    "kind": "开源",
    "src": "https://huggingface.co/sesame/csm-1b"
   },
   "total_raised": "累计融资约 $3.08 亿（Forge 口径 $307.62M）",
   "funding": [
    {
     "date": "2025-10",
     "round": "B 轮",
     "amount": "约 $2.5 亿",
     "lead": "Sequoia、Spark Capital（Sequoia 官方称联合领投）",
     "src": "https://sequoiacap.com/article/partnering-with-sesame-a-new-era-for-voice"
    }
   ],
   "key_data": [
    {
     "k": "Hugging Face 近 30 天下载（csm-1b）",
     "v": "102,269",
     "date": "2026-11",
     "src": "https://huggingface.co/sesame/csm-1b"
    },
    {
     "k": "累计融资 Forge 页",
     "v": "$307.62M",
     "date": "2026-10",
     "src": "https://forgeglobal.com/sesame-ai_stock/"
    }
   ],
   "note": "拿不准是否算端侧：Sequoia 官方页没提端侧，只有 CSM-1B 开源权重可本地跑，产品主体在云端；眼镜是在研硬件。建议交回由你决定是否保留。B 轮金额取 Tracxn / Forge 的 $250M / $248M，Sequoia 页未写金额。估值各源冲突（$8.9 亿到 $11.9 亿不等），不采。创立年份取百度百科（2023，低权威）。"
  },
  {
   "name": "Cerence",
   "layer": "m5",
   "url": "https://www.cerence.com",
   "hq": "伯灵顿",
   "region": "海外",
   "type": "创业",
   "founded": 2019,
   "desc": "车载语音 AI 上市公司，CaLLM Edge 小语言模型在车机芯片上本地运行、断网可用",
   "modality": "语音",
   "products": [
    "CaLLM Edge（车载端侧 SLM）",
    "Cerence xUI"
   ],
   "devices": [
    "车机"
   ],
   "chips": [
    "Qualcomm Hexagon",
    "MediaTek APU"
   ],
   "license": {
    "v": "商业闭源",
    "kind": "闭源",
    "src": "https://www.cerence.com"
   },
   "revenue": {
    "v": "FY2025（截至 2025-09-30）营收 $2.518 亿",
    "date": "2025-09",
    "kind": "财报，公司公布",
    "src": "https://www.cerence.cn/newsinfo/10828378.html"
   },
   "total_raised": "融资不适用（Nasdaq: CRNC）",
   "key_data": [
    {
     "k": "搭载 Cerence 技术的全球汽车产量占比（TTM）",
     "v": "52%",
     "date": "2025-09",
     "src": "https://www.cerence.cn/newsinfo/10828378.html"
    },
    {
     "k": "5 年积压订单",
     "v": "$10 亿",
     "date": "2025-09",
     "src": "https://www.cerence.cn/newsinfo/10828378.html"
    },
    {
     "k": "CaLLM Edge 通常跑在中档 Qualcomm / MediaTek / ARM SoC，需 4–7GB 内存（公司高管在 Needham 会议所述）",
     "v": "4–7GB",
     "date": "2026-01",
     "src": "https://stockanalysis.com/stocks/crnc/transcripts/537706-28th-annual-needham-growth-conference-virtual/"
    }
   ],
   "public": true,
   "note": "founded 取 2019（自 Nuance 分拆），本轮未查到一手出处。valueray 页写 FY2025 营收 $4.21 亿，与公司公告 $2.518 亿冲突，不采。CaLLM Edge 与 Arm、SiMa.ai 合作的说法只见 valueray，未采。"
  },
  {
   "name": "Kyutai",
   "layer": "m5",
   "url": "https://kyutai.org",
   "hq": "巴黎",
   "region": "海外",
   "type": "研究机构",
   "founded": 2023,
   "founded_month": "2023-11",
   "desc": "非营利开源语音实验室：全双工对话模型 Moshi、流式 STT / TTS 以及 100M 的 CPU 实时 Pocket TTS",
   "modality": "语音",
   "products": [
    "Moshi（全双工）",
    "Unmute（STT + TTS）",
    "Pocket TTS 100M",
    "Hibiki 同传"
   ],
   "params": {
    "v": "Moshi 7B；Pocket TTS 100M；STT 1B / 2.6B",
    "src": "https://kyutai.org/"
   },
   "devices": [
    "Mac",
    "Linux"
   ],
   "formats": [
    "MLX"
   ],
   "license": {
    "v": "权重 CC-BY-4.0；代码 MIT / Apache-2.0",
    "kind": "开源",
    "src": "https://huggingface.co/kyutai/pocket-tts"
   },
   "benchmarks": [
    {
     "k": "Open ASR 平均 WER（kyutai/stt-2.6b-en，第 38）",
     "v": "6.78%",
     "date": "2026-08",
     "src": "https://huggingface.co/spaces/hf-audio/open_asr_leaderboard"
    }
   ],
   "total_raised": "非营利，三位出资人合计约 €3 亿",
   "key_data": [
    {
     "k": "Hugging Face 近 30 天下载（moshiko-pytorch-bf16）",
     "v": "201,081",
     "date": "2026-11",
     "src": "https://huggingface.co/kyutai/moshiko-pytorch-bf16"
    },
    {
     "k": "Pocket TTS 在 CPU 上实时运行",
     "v": "官方描述",
     "date": "2026-11",
     "src": "https://kyutai.org/"
    },
    {
     "k": "启动资金",
     "v": "€3 亿，2023-11-17 启动",
     "date": "2023-11",
     "src": "https://converteo.com/ressources/kyutai-laboratoire-ia-patrick-perez-moshi/"
    }
   ],
   "note": "Gradium 是其首个分拆公司（m4）。Moshi 7B 与 Apple Silicon（MLX）运行来自 AI Tool Bible 页（低权威），Moshi 手机端部署未查到。出资人姓名只出现在引用页里，卡内不当主角。"
  },
  {
   "name": "openWakeWord",
   "layer": "m5",
   "url": "https://github.com/dscripka/openWakeWord",
   "region": "海外",
   "type": "开源组织",
   "desc": "开源唤醒词库，树莓派上实时跑，Home Assistant 语音生态常用",
   "modality": "语音",
   "products": [
    "openWakeWord"
   ],
   "devices": [
    "树莓派",
    "Linux"
   ],
   "license": {
    "v": "代码 Apache-2.0；官方预训练模型 CC BY-NC-SA 4.0（非商用）",
    "kind": "开源",
    "src": "https://huggingface.co/davidscripka/openwakeword"
   },
   "total_raised": "无公司主体，融资不适用",
   "note": "个人项目，hq、region、founded 省略（未查）。Home Assistant 生态说法来自初稿，本轮未单独核；树莓派运行为推断。"
  },
  {
   "name": "Sensory",
   "layer": "m5",
   "url": "https://www.sensory.com",
   "hq": "圣克拉拉",
   "region": "海外",
   "type": "创业",
   "founded": 1994,
   "desc": "嵌入式唤醒词与离线语音的老牌公司，TrulyHandsfree 跑在 MCU / DSP 上，2025-12 推 Smart Wakewords",
   "modality": "语音",
   "products": [
    "TrulyHandsfree",
    "Smart Wakewords（2025-12）",
    "TrulyNatural"
   ],
   "devices": [
    "MCU",
    "车机"
   ],
   "license": {
    "v": "商业授权",
    "kind": "闭源",
    "src": "https://www.sensory.com"
   },
   "total_raised": "累计融资未披露（PitchBook：私营，曾有风投；员工约 60 人）",
   "key_data": [
    {
     "k": "TrulyHandsfree 接入 Ceva NeuPro-Nano NPU",
     "v": "2026-01-05 联合新闻稿",
     "date": "2026-01",
     "src": "https://www.nasdaq.com/press-release/ceva-enhances-neupro-nano-npu-ecosystem-addition-sensorys-trulyhandsfree-voice"
    }
   ],
   "customers": [
    "Ceva（NeuPro-Nano NPU 生态合作）",
    "Telly（智能电视）"
   ],
   "note": "官网对脚本返回 403，已用法院文件（成立于 1994 年圣克拉拉）和 Nasdaq 新闻稿交叉印证：https://storage.courtlistener.com/recap/gov.uscourts.dcd.273293/gov.uscourts.dcd.273293.18.0.pdf 。Telly 来自 Owler 汇总（低权威）。基准未查到。"
  },
  {
   "name": "SoundHound AI",
   "layer": "m5",
   "url": "https://www.soundhound.com",
   "hq": "圣克拉拉",
   "region": "海外",
   "type": "创业",
   "desc": "语音 AI 上市公司，2026-09 发 OASYS Edge：LLM 语音智能体直接跑在车机和设备上，可离线",
   "modality": "语音",
   "products": [
    "OASYS Edge（2026-09）",
    "Houndify 嵌入式",
    "SoundHound Chat AI"
   ],
   "devices": [
    "车机"
   ],
   "license": {
    "v": "商业闭源",
    "kind": "闭源",
    "src": "https://www.soundhound.com"
   },
   "revenue": {
    "v": "2026 年 Q2 营收 $6,190 万，同比 +45%",
    "date": "2026-Q2",
    "kind": "财报，公司公布",
    "src": "https://investors.soundhound.com/node/11241/pdf"
   },
   "total_raised": "融资不适用（Nasdaq: SOUN）",
   "key_data": [
    {
     "k": "OASYS Edge 发布",
     "v": "2026-09-24 新闻稿",
     "date": "2026-09",
     "src": "https://investors.soundhound.com/news-releases/news-release-details/soundhound-ai-introduces-oasys-edge-bringing-fully-embedded"
    },
    {
     "k": "GTC 2026 展示完全在端侧运行的多模态智能体",
     "v": "2026-03-16",
     "date": "2026-03",
     "src": "https://www.nasdaq.com/press-release/soundhound-ai-unveils-worlds-first-multimodal-agentic-ai-completely-edge-nvidia-gtc"
    }
   ],
   "public": true,
   "note": "主线含云端语音与餐饮 / 客服等业务，车载和设备端嵌入式是其一块，收录依据是 OASYS Edge。FY2025 营收 $1.689 亿（+99%）见 stockanalysis / saasvaluation 汇总（低权威），未用。成立年份、市值未核。"
  },
  {
   "name": "云知声 Unisound",
   "layer": "m5",
   "url": "https://www.unisound.com",
   "hq": "北京",
   "region": "中国",
   "type": "创业",
   "founded": 2012,
   "desc": "港股上市的语音与大模型公司，自研 AI 芯片出货破 1.1 亿颗，做家居、车载等离线语音",
   "modality": "语音",
   "products": [
    "山海大模型",
    "AI 语音芯片",
    "离线语音"
   ],
   "devices": [
    "MCU",
    "车机"
   ],
   "license": {
    "v": "商业授权",
    "kind": "闭源",
    "src": "https://www.unisound.com"
   },
   "valuation": {
    "v": "上市首日收盘市值 221.37 亿港元",
    "date": "2025-06-30",
    "kind": "上市日市值，媒体转述",
    "src": "https://www.c114.net.cn/industry/71340.html"
   },
   "revenue": {
    "v": "2025 年营收 12.1 亿元，同比 +29%",
    "date": "2025",
    "kind": "年报，公司公布",
    "src": "https://www.unisound.com/en/news/139"
   },
   "total_raised": "已于 2025-06-30 在港交所上市（09678.HK）",
   "key_data": [
    {
     "k": "AI 芯片出货量",
     "v": "破 1.1 亿颗",
     "date": "2025",
     "src": "https://www.c114.net.cn/industry/71340.html"
    }
   ],
   "public": true,
   "note": "补充收录（初稿漏）。founded 2012、总部北京为背景认知，未在引用页验证。具体离线语音产品页本轮未打开，收录依据是 1.1 亿颗 AI 芯片出货及语音业务；大模型收入占比已超一半（2025 年 6.1 亿元，公司公告），端侧属性偏弱，建议你决定去留。"
  },
  {
   "name": "思必驰 AISpeech",
   "layer": "m5",
   "url": "https://www.aispeech.com",
   "hq": "苏州",
   "region": "中国",
   "type": "创业",
   "founded": 2007,
   "desc": "对话式 AI 公司，车载语音装机量位居行业第二，自研 AI 芯片出货过千万、离线语音上车",
   "modality": "语音",
   "products": [
    "车载语音（智慧出行）",
    "自研 AI 芯片",
    "AI 办公本"
   ],
   "devices": [
    "车机"
   ],
   "license": {
    "v": "商业授权",
    "kind": "闭源",
    "src": "https://www.aispeech.com"
   },
   "valuation": {
    "v": "拟冲击估值约 64 亿元",
    "date": "2026-05",
    "kind": "媒体报道（IPO 申报期）",
    "src": "https://www.sohu.com/a/1030006802_122014422"
   },
   "revenue": {
    "v": "2025 年营收 6.88 亿元，同比 +14.5%",
    "date": "2025",
    "kind": "招股书披露（三个皮匠转述）",
    "src": "https://www.sgpjbg.com/labelsyh/sibichi2026zhaogushujiedu/1/7625740.html"
   },
   "total_raised": "累计融资未取到一手汇总",
   "key_data": [
    {
     "k": "车载语音装机量市占率（盖世汽车研究院，2025，招股书引用）",
     "v": "22%，行业第二",
     "date": "2025",
     "src": "https://www.sgpjbg.com/labelsyh/sibichifazhanqianjing/1/7625740.html"
    },
    {
     "k": "累计上车",
     "v": "超 2,500 万辆",
     "date": "2026-05",
     "src": "https://www.sgpjbg.com/labelsyh/sibichifazhanqianjing/1/7625740.html"
    },
    {
     "k": "IPO：2026-05-25 再次向上交所递交科创板招股书，拟募资 15.55 亿元（2023-05 曾终止）",
     "v": "申报中",
     "date": "2026-05",
     "src": "https://lieyunpro.com/archives/501463"
    }
   ],
   "customers": [
    "比亚迪",
    "上汽",
    "吉利",
    "长城",
    "赛力斯",
    "梅赛德斯-奔驰",
    "奥迪",
    "大众"
   ],
   "note": "更正初稿：思必驰目前在申请科创板 IPO，不是港股上市公司，故 pub 不填。招股书原文未直接打开，数字取自转述文章（三个皮匠报告、猎云网，权威偏低），建议二次用上交所披露页核对。端侧产品具体型号（离线 / 本地大模型）本轮未查到，收录依据是车载嵌入式语音与自研芯片。"
  },
  {
   "name": "阿里 Qwen-Omni 全模态语音模型",
   "layer": "m5",
   "url": "https://qwen.ai",
   "hq": "杭州",
   "region": "中国",
   "type": "大厂产品线",
   "founded": 2025,
   "desc": "千问的端到端语音输入输出模型，3B / 7B 小尺寸可上手机和 PC，30B-A3B 面向服务器和高端端侧",
   "modality": "多模态",
   "products": [
    "Qwen2.5-Omni 3B / 7B",
    "Qwen3-Omni-30B-A3B"
   ],
   "params": {
    "v": "3B / 7B / 30B-A3B（总参）",
    "src": "https://huggingface.co/Qwen/Qwen2.5-Omni-3B"
   },
   "devices": [
    "Android",
    "Mac",
    "Windows",
    "Linux"
   ],
   "license": {
    "v": "Qwen3-Omni Apache-2.0；Qwen2.5-Omni-3B Qwen 研究许可",
    "kind": "开放权重",
    "src": "https://huggingface.co/Qwen/Qwen2.5-Omni-3B"
   },
   "total_raised": "融资不适用（阿里巴巴）",
   "key_data": [
    {
     "k": "Hugging Face 近 30 天下载（Qwen3-Omni-30B-A3B-Instruct）",
     "v": "546,836",
     "date": "2026-11",
     "src": "https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct"
    },
    {
     "k": "Hugging Face 近 30 天下载（Qwen2.5-Omni-3B）",
     "v": "96,889",
     "date": "2026-11",
     "src": "https://huggingface.co/Qwen/Qwen2.5-Omni-3B"
    }
   ],
   "product_line_of": "阿里巴巴",
   "public": true,
   "note": "与 m1 的 Qwen 小尺寸同属一家，这里为语音维度单列。30B-A3B 总参超 10B。是否为全双工（可打断）未核。founded 取 Qwen2.5-Omni-3B 最后更新 2025-04（HF），首发日期未核。基准未查到。"
  },
  {
   "name": "GPT4All（Nomic）",
   "layer": "s1",
   "url": "https://www.nomic.ai/gpt4all",
   "hq": "纽约",
   "region": "海外",
   "type": "创业",
   "founded": 2022,
   "desc": "本地聊天桌面应用，出自 Nomic",
   "modality": "通用",
   "products": [
    "GPT4All",
    "Nomic Embed"
   ],
   "devices": [
    "Mac",
    "Windows",
    "Linux"
   ],
   "chips": [
    "Apple GPU (Metal)",
    "NVIDIA GPU",
    "AMD GPU"
   ],
   "formats": [
    "GGUF"
   ],
   "license": {
    "v": "MIT",
    "kind": "开源",
    "src": "https://github.com/nomic-ai/gpt4all/blob/main/LICENSE.txt"
   },
   "valuation": {
    "v": "约 $1 亿",
    "date": "2023-07",
    "kind": "A 轮投后，媒体报道",
    "src": "https://siliconangle.com/2023/07/13/ai-startup-nomic-raises-17m-build-open-source-alternative-gpt-4/"
   },
   "total_raised": "累计融资约 $19M（聚合页口径，含早期种子 $2M；仅 A 轮有一手出处）",
   "funding": [
    {
     "date": "2023-07",
     "round": "A 轮",
     "amount": "$17M",
     "lead": "Coatue（Contrary Capital、Betaworks Ventures、SV Angel、Story Ventures、Factorial Capital 跟投）",
     "post": "投后约 $1 亿",
     "src": "https://siliconangle.com/2023/07/13/ai-startup-nomic-raises-17m-build-open-source-alternative-gpt-4/"
    }
   ],
   "key_data": [
    {
     "k": "GitHub stars",
     "v": "77,348",
     "date": "2026-11-03",
     "src": "https://github.com/nomic-ai/gpt4all"
    }
   ],
   "note": "GPT4All 仓库仍高星，但 2026 年二手报道称公司主线已转向工程图纸方向（KuCoin 2026-09-03 转述，低权威，未核），交调研方判断是否保留。无基准。"
  },
  {
   "name": "Jan（Menlo Research）",
   "layer": "s1",
   "url": "https://jan.ai",
   "hq": "新加坡",
   "region": "海外",
   "type": "创业",
   "founded": 2023,
   "founded_month": "2023-10",
   "desc": "开源本地 ChatGPT 替代，桌面 App，带 MLX",
   "modality": "通用",
   "products": [
    "Jan",
    "Jan-nano",
    "Jan-v1"
   ],
   "devices": [
    "Mac",
    "Windows",
    "Linux"
   ],
   "chips": [
    "Apple GPU (Metal)",
    "NVIDIA GPU"
   ],
   "formats": [
    "GGUF",
    "MLX"
   ],
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://github.com/janhq/jan/blob/main/LICENSE"
   },
   "total_raised": "累计融资未披露（Tech in Asia 2025-05 报道称未拿 VC 融资）",
   "key_data": [
    {
     "k": "GitHub stars",
     "v": "45,404",
     "date": "2026-11-03",
     "src": "https://github.com/janhq/jan"
    }
   ],
   "note": "Menlo Research Pte. Ltd.（原 Homebrew，2025-03 改名，称做「机器人的大脑」）。2025-05 起协议为 Apache-2.0（此前 AGPL，二手来源）。无基准。融资依据 https://www.linkedin.com/posts/tech-in-asia_this-startup-builds-robot-brains-with-open-source-activity-7334191551276883968-U3tS ；改名见 https://menlo.ai/research/homebrew-to-menlo"
  },
  {
   "name": "LM Studio",
   "layer": "s1",
   "url": "https://lmstudio.ai",
   "hq": "纽约",
   "region": "海外",
   "type": "创业",
   "founded": 2023,
   "founded_month": "2023-05",
   "desc": "桌面本地模型应用，含 OpenAI 兼容服务和 MLX",
   "modality": "通用",
   "products": [
    "LM Studio",
    "lms CLI",
    "LM Studio SDK",
    "Bionic"
   ],
   "devices": [
    "Mac",
    "Windows",
    "Linux"
   ],
   "chips": [
    "Apple GPU (Metal)",
    "NVIDIA GPU",
    "AMD GPU"
   ],
   "formats": [
    "GGUF",
    "MLX"
   ],
   "license": {
    "v": "应用闭源；mlx-engine 为 MIT",
    "kind": "闭源",
    "src": "https://faun.dev/company/element-labs/"
   },
   "total_raised": "累计融资约 $19.3M（据 SEC Form D，转述页；公司未发通稿）",
   "funding": [
    {
     "date": "2025-04",
     "round": "早期股权融资（SEC Form D）",
     "amount": "约 $19.3M（含约 $3.97M 可转债转换）",
     "lead": "领投方未见官方披露",
     "src": "https://wenjiai.cn/products/lm-studio"
    }
   ],
   "key_data": [
    {
     "k": "GitHub stars（lms CLI 仓库）",
     "v": "5,404",
     "date": "2026-11-03",
     "src": "https://github.com/lmstudio-ai/lms"
    },
    {
     "k": "收购",
     "v": "2026-04 收购 Locally AI（iOS 本地聊天应用）",
     "date": "2026-04",
     "src": "https://faun.dev/company/element-labs/"
    }
   ],
   "note": "运营公司 Element Labs Inc.（纽约布鲁克林）。不要与以色列同名 Element Labs（Globes 报道 $4B 估值）混淆。数据库把 Matrix、Torch Capital、Preston-Werner Ventures 列为投资方，Torch 在 LinkedIn 自称投资人，领投未核。桌面 App 2025-07 起免费商用。无基准。"
  },
  {
   "name": "Mozilla llamafile",
   "layer": "s1",
   "url": "https://github.com/mozilla-ai/llamafile",
   "region": "海外",
   "type": "开源组织",
   "founded": 2023,
   "desc": "把模型和运行时打成单个可执行文件",
   "modality": "通用",
   "products": [
    "llamafile"
   ],
   "devices": [
    "Mac",
    "Windows",
    "Linux"
   ],
   "chips": [
    "Apple GPU (Metal)",
    "NVIDIA GPU",
    "AMD GPU"
   ],
   "formats": [
    "GGUF"
   ],
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://github.com/mozilla-ai/llamafile/blob/main/LICENSE"
   },
   "key_data": [
    {
     "k": "GitHub stars",
     "v": "26,555",
     "date": "2026-11-03",
     "src": "https://github.com/mozilla-ai/llamafile"
    }
   ],
   "note": "Mozilla.ai 维护（Mozilla 旗下）。总部城市、融资本轮未查。无基准。"
  },
  {
   "name": "Ollama",
   "layer": "s1",
   "url": "https://ollama.com",
   "hq": "帕洛阿尔托",
   "region": "海外",
   "type": "创业",
   "founded": 2023,
   "desc": "一行命令在本机跑开源模型，另有云端托管",
   "modality": "通用",
   "products": [
    "Ollama 桌面/CLI",
    "Ollama Cloud"
   ],
   "devices": [
    "Mac",
    "Windows",
    "Linux"
   ],
   "chips": [
    "Apple GPU (Metal)",
    "NVIDIA GPU",
    "AMD GPU",
    "Qualcomm Hexagon",
    "Intel NPU"
   ],
   "formats": [
    "GGUF"
   ],
   "license": {
    "v": "MIT",
    "kind": "开源",
    "src": "https://github.com/ollama/ollama/blob/main/LICENSE"
   },
   "total_raised": "累计融资约 $88M（公司口径，含此前 A 轮 $15M，Benchmark 领投）",
   "funding": [
    {
     "date": "2026-07",
     "round": "B 轮",
     "amount": "$65M",
     "lead": "Theory Ventures（Benchmark、8VC、Y Combinator、Pace Capital、49 Palms、GTMfund 跟投）",
     "src": "https://techcrunch.com/2026/07/09/popular-open-source-ai-developer-tool-ollama-raises-65m-grows-to-nearly-9m-users/"
    }
   ],
   "key_data": [
    {
     "k": "GitHub stars",
     "v": "185,818",
     "date": "2026-11-03",
     "src": "https://github.com/ollama/ollama"
    },
    {
     "k": "月活开发者",
     "v": "约 890 万（创始人口径，未经第三方拆分）",
     "date": "2026-07",
     "src": "https://thenextweb.com/news/ollama-65m-series-b-theory-ventures-open-models"
    }
   ],
   "customers": [
    "Meta",
    "Google DeepMind",
    "Mistral",
    "NVIDIA",
    "Intel",
    "AMD",
    "Qualcomm"
   ],
   "note": "估值、营收未披露。YC W21 公司，产品 2023 年公开。customers 为模型/芯片发行合作方（TNW 2026-07-08）。无官方 tokens/s 基准。"
  },
  {
   "name": "Cactus Compute",
   "layer": "s2",
   "url": "https://www.cactuscompute.com",
   "hq": "旧金山",
   "region": "海外",
   "type": "创业",
   "founded": 2025,
   "desc": "手机与可穿戴专用推理引擎，带混合云回退",
   "modality": "通用",
   "products": [
    "Cactus engine",
    "Needle"
   ],
   "devices": [
    "iPhone",
    "Android",
    "Mac",
    "MCU"
   ],
   "chips": [
    "Apple ANE",
    "Qualcomm Hexagon",
    "Arm CPU"
   ],
   "formats": [
    "GGUF",
    "Cactus 量化格式"
   ],
   "license": {
    "v": "自定义许可：个人/教育/非营利免费；融资与年营收均低于 $200 万的组织免费",
    "kind": "开源",
    "src": "https://github.com/cactus-compute/cactus/blob/main/LICENSE"
   },
   "total_raised": "累计融资未披露",
   "funding": [
    {
     "date": "2025",
     "round": "种子轮",
     "amount": "金额未披露",
     "lead": "FundersClub、Oxford Seed Fund、Y Combinator（S25）",
     "src": "https://www.linkedin.com/posts/activity-7413349550205071360-Mgan"
    },
    {
     "date": "2026-08",
     "round": "战略投资（一家大型科技公司）",
     "amount": "金额、投资方未披露",
     "lead": "未披露",
     "src": "https://www.linkedin.com/posts/henry-ndubuaku-7b6350b8_biggest-lessons-as-an-early-stage-founder-activity-7492965892805382144-23JW"
    }
   ],
   "key_data": [
    {
     "k": "GitHub stars",
     "v": "6,173",
     "date": "2026-11-03",
     "src": "https://github.com/cactus-compute/cactus"
    }
   ],
   "note": "源码公开但许可非 OSI 开源（有收入与融资门槛），license.kind 按「开源」归类，建站时可改称「源码公开」。融资两条均为创始人 LinkedIn 自述。聚合站写 $7M 种子，无一手出处，不采。无基准。"
  },
  {
   "name": "ggml.ai（llama.cpp / ggml）",
   "layer": "s2",
   "url": "https://ggml.ai",
   "region": "海外",
   "type": "开源组织",
   "founded": 2023,
   "founded_month": "2023-03",
   "desc": "llama.cpp 与 GGUF 格式，端侧推理事实标准",
   "modality": "通用",
   "products": [
    "llama.cpp",
    "ggml",
    "whisper.cpp",
    "GGUF"
   ],
   "devices": [
    "iPhone",
    "Android",
    "Mac",
    "Windows",
    "Linux",
    "树莓派",
    "浏览器"
   ],
   "chips": [
    "Apple GPU (Metal)",
    "NVIDIA GPU",
    "AMD GPU",
    "Qualcomm Hexagon",
    "Arm CPU (KleidiAI)"
   ],
   "formats": [
    "GGUF"
   ],
   "license": {
    "v": "MIT",
    "kind": "开源",
    "src": "https://github.com/ggml-org/llama.cpp/blob/master/LICENSE"
   },
   "benchmarks": [
    {
     "k": "M4 Max · Nemotron-3 Nano 30B-A3B Q4_K_M 解码",
     "v": "86.19 tok/s（llama.cpp b8680，Metal，第三方实测）",
     "date": "2026-09",
     "src": "https://github.com/john-rocky/apple-silicon-llm-bench/blob/bd5feaf9fb1a5539046bb30789223d628d39d622/results/hybrid/nemotron3-nano-apple-silicon-bench.md"
    },
    {
     "k": "M4 Max · Gemma-4-E2B Q4_K_S 解码（Metal，第三方实测）",
     "v": "140.48 tok/s",
     "date": "2026-07",
     "src": "https://github.com/leehack/llamadart/blob/a395ebdd5a8c3076fd8d1cd746559b59094d48cf/website/versioned_docs/version-0.8.13/guides/backend-benchmarks.md"
    }
   ],
   "total_raised": "累计融资未披露",
   "funding": [
    {
     "date": "2023",
     "round": "Pre-seed",
     "amount": "金额未披露",
     "lead": "Nat Friedman、Daniel Gross",
     "src": "https://ggml.ai/"
    }
   ],
   "key_data": [
    {
     "k": "GitHub stars（llama.cpp）",
     "v": "133,295",
     "date": "2026-11-03",
     "src": "https://github.com/ggml-org/llama.cpp"
    },
    {
     "k": "GitHub stars（whisper.cpp）",
     "v": "54,895",
     "date": "2026-11-03",
     "src": "https://github.com/ggml-org/whisper.cpp"
    },
    {
     "k": "GitHub stars（ggml）",
     "v": "15,580",
     "date": "2026-11-03",
     "src": "https://github.com/ggml-org/ggml"
    }
   ],
   "owner": "2026-02-20 ggml.ai 并入 Hugging Face（金额未披露）；项目继续开源、由原团队全职维护",
   "note": "并入 Hugging Face 已查实：HF 官方博客 https://huggingface.co/blog/ggml-joins-hf 、GitHub 公告 https://github.com/ggml-org/llama.cpp/discussions/19759 、ggml.ai 首页自述。总部城市未查到一手出处。基准均为第三方实测，非官方。"
  },
  {
   "name": "KTransformers（清华 MADSys / 趋境）",
   "layer": "s2",
   "url": "https://github.com/kvcache-ai/ktransformers",
   "hq": "北京",
   "region": "中国",
   "type": "研究机构",
   "founded": 2024,
   "desc": "CPU+GPU 异构推理，本地跑大型 MoE 模型",
   "modality": "通用",
   "products": [
    "KTransformers"
   ],
   "devices": [
    "Linux",
    "Windows"
   ],
   "chips": [
    "NVIDIA GPU",
    "Intel CPU (AMX)",
    "AMD GPU"
   ],
   "formats": [
    "GGUF",
    "FP8/INT4 权重"
   ],
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://github.com/kvcache-ai/ktransformers/blob/main/LICENSE"
   },
   "key_data": [
    {
     "k": "GitHub stars",
     "v": "19,659",
     "date": "2026-11-03",
     "src": "https://github.com/kvcache-ai/ktransformers"
    }
   ],
   "note": "面向桌面/工作站「本地」而非手机。2025-10 起接入 SGLang（README）。商业公司趋境科技融资本轮未查。README 吞吐表列与模型对应不清，未取数。founded 估写，请核。"
  },
  {
   "name": "Microsoft ONNX Runtime",
   "layer": "s2",
   "url": "https://onnxruntime.ai",
   "hq": "雷德蒙德",
   "region": "海外",
   "type": "大厂产品线",
   "founded": 2018,
   "desc": "跨平台推理，GenAI 扩展跑 Phi 等小模型",
   "modality": "通用",
   "products": [
    "ONNX Runtime",
    "ONNX Runtime GenAI"
   ],
   "devices": [
    "Windows",
    "Android",
    "iPhone",
    "Mac",
    "Linux",
    "浏览器"
   ],
   "chips": [
    "Qualcomm Hexagon (QNN)",
    "Intel NPU (OpenVINO)",
    "AMD XDNA (Vitis AI)",
    "NVIDIA GPU",
    "Apple ANE (Core ML)",
    "DirectML"
   ],
   "formats": [
    "ONNX",
    "ORT"
   ],
   "license": {
    "v": "MIT",
    "kind": "开源",
    "src": "https://github.com/microsoft/onnxruntime/blob/main/LICENSE"
   },
   "key_data": [
    {
     "k": "GitHub stars（onnxruntime）",
     "v": "22,218",
     "date": "2026-11-03",
     "src": "https://github.com/microsoft/onnxruntime"
    },
    {
     "k": "GitHub stars（onnxruntime-genai）",
     "v": "1,149",
     "date": "2026-11-03",
     "src": "https://github.com/microsoft/onnxruntime-genai"
    }
   ],
   "product_line_of": "Microsoft",
   "public": true,
   "note": "首次开源年份 2018 凭背景知识，未逐项核。无官方统一 tokens/s 基准，省略。"
  },
  {
   "name": "MLC LLM / WebLLM",
   "layer": "s2",
   "url": "https://llm.mlc.ai",
   "region": "海外",
   "type": "开源组织",
   "founded": 2023,
   "desc": "TVM 编译，手机和浏览器里跑 LLM",
   "modality": "通用",
   "products": [
    "MLC LLM",
    "WebLLM"
   ],
   "devices": [
    "iPhone",
    "Android",
    "Mac",
    "Windows",
    "Linux",
    "浏览器"
   ],
   "chips": [
    "Apple GPU (Metal)",
    "Qualcomm Adreno (OpenCL)",
    "NVIDIA GPU",
    "AMD GPU",
    "WebGPU"
   ],
   "formats": [
    "MLC 编译权重"
   ],
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://github.com/mlc-ai/mlc-llm/blob/main/LICENSE"
   },
   "key_data": [
    {
     "k": "GitHub stars",
     "v": "23,301",
     "date": "2026-11-03",
     "src": "https://github.com/mlc-ai/mlc-llm"
    }
   ],
   "note": "项目由 MLC.ai 社区维护，公司主体、总部、融资本轮未查到一手出处，不写。初稿里「OctoAI 被 NVIDIA 收购」未核，不采。founded 凭背景知识，未核。无官方基准。"
  },
  {
   "name": "mllm",
   "layer": "s2",
   "url": "https://github.com/UbiquitousLearning/mllm",
   "region": "中国",
   "type": "研究机构",
   "founded": 2023,
   "desc": "面向手机与边缘 NPU 的轻量推理引擎",
   "modality": "通用",
   "products": [
    "mllm",
    "pymllm"
   ],
   "devices": [
    "Android",
    "Jetson",
    "Linux"
   ],
   "chips": [
    "Qualcomm Hexagon (QNN AOT)",
    "Ascend NPU",
    "NVIDIA GPU (Jetson Orin)",
    "Arm CPU"
   ],
   "formats": [
    "mllm 格式"
   ],
   "license": {
    "v": "MIT",
    "kind": "开源",
    "src": "https://github.com/UbiquitousLearning/mllm/blob/main/LICENSE"
   },
   "benchmarks": [
    {
     "k": "Jetson AGX Orin 32GB · Qwen3-VL-2B W8A8 prefill（输入 2048 / 输出 128）",
     "v": "最高为 llama.cpp 的 3.12 倍，约 12,243 tok/s；解码与 llama.cpp 接近（README 自述）",
     "date": "2026-06",
     "src": "https://github.com/UbiquitousLearning/mllm"
    }
   ],
   "key_data": [
    {
     "k": "GitHub stars",
     "v": "1,633",
     "date": "2026-11-03",
     "src": "https://github.com/UbiquitousLearning/mllm"
    }
   ],
   "note": "所属单位、总部城市、founded 年份本轮未查到一手出处，founded 为估写，请核。"
  },
  {
   "name": "Nexa AI（NexaSDK / GenieX）",
   "layer": "s2",
   "url": "https://github.com/qualcomm/GenieX",
   "region": "海外",
   "type": "创业",
   "founded": 2023,
   "desc": "NPU 优先的跨芯片推理 SDK，现为高通 GenieX",
   "modality": "通用",
   "products": [
    "NexaSDK",
    "GenieX",
    "Hyperlink",
    "Octopus"
   ],
   "devices": [
    "Windows",
    "Android",
    "Linux",
    "Mac"
   ],
   "chips": [
    "Qualcomm Hexagon",
    "Apple ANE",
    "Intel NPU",
    "AMD XDNA",
    "NVIDIA GPU"
   ],
   "formats": [
    "GGUF",
    "QNN",
    "MLX"
   ],
   "license": {
    "v": "BSD-3-Clause（GenieX）",
    "kind": "开源",
    "src": "https://github.com/qualcomm/GenieX/blob/main/LICENSE"
   },
   "total_raised": "累计融资未披露（PitchBook 页列 $16.5M，仅线索，未核）",
   "funding": [
    {
     "date": "2024",
     "round": "早期融资",
     "amount": "金额未披露（2024 年访谈转述「种子轮超千万美元」，低权威）",
     "lead": "Fusion Fund",
     "src": "https://www.linkedin.com/posts/luzhangvc_ondeviceai-edgeai-agenticai-activity-7480321547346305024-ifBO"
    }
   ],
   "key_data": [
    {
     "k": "GitHub stars（原 NexaAI/nexa-sdk，已迁至 qualcomm/GenieX）",
     "v": "8,490",
     "date": "2026-11-03",
     "src": "https://github.com/qualcomm/GenieX"
    }
   ],
   "owner": "被 Qualcomm 收购（金额未披露）：12 人团队 2026-03 中旬加入；公开官宣 2026-07-06（Nexa 与 Fusion Fund 同日发帖）；2026-07 以 GenieX 开发者预览版发布",
   "note": "收购时间已解冲突：初稿「2025-03」无任何出处，不采。依据：Qualcomm 官方博文（CSDN 转载 2026-07-22）写「今年早些时候收购」；硅星人访谈（品玩 2026-07-15）写 2026 年 3 月中旬团队加入；PitchBook 记并购日期 2026-03-13（线索）；Nexa LinkedIn 与 Fusion Fund 帖均为 2026-07-06。https://www.pingwest.com/a/315521 ；https://tianqi.csdn.net/6a6179e4662f9a54cb936614.html 。总部城市未核。无基准。"
  },
  {
   "name": "PowerInfer / Tiiny AI",
   "layer": "s2",
   "url": "https://github.com/Tiiny-AI/PowerInfer",
   "region": "中国",
   "type": "创业",
   "founded": 2024,
   "desc": "稀疏激活推理引擎，出自上交 IPADS 实验室",
   "modality": "通用",
   "products": [
    "PowerInfer",
    "PowerInfer-2",
    "Tiiny AI Pocket Lab"
   ],
   "devices": [
    "Android",
    "Mac",
    "Linux",
    "Windows"
   ],
   "chips": [
    "NVIDIA GPU",
    "Qualcomm Snapdragon 8 Gen 3",
    "Arm CPU"
   ],
   "formats": [
    "GGUF（自有转换脚本）"
   ],
   "license": {
    "v": "MIT",
    "kind": "开源",
    "src": "https://github.com/Tiiny-AI/PowerInfer/blob/main/LICENSE"
   },
   "benchmarks": [
    {
     "k": "PowerInfer · RTX 4090 · Falcon-40B ReLU FP16 对 llama.cpp",
     "v": "最高 11x（README 自述）",
     "date": "2024",
     "src": "https://github.com/Tiiny-AI/PowerInfer"
    },
    {
     "k": "PowerInfer-2 · 手机 · TurboSparse-Mixtral-47B",
     "v": "11.68 tok/s（官方自述）",
     "date": "2024-06",
     "src": "https://arxiv.org/abs/2406.06282"
    }
   ],
   "total_raised": "累计融资未披露（仅此一轮有媒体出处）",
   "funding": [
    {
     "date": "2025",
     "round": "种子轮",
     "amount": "数千万人民币",
     "lead": "光启资本（BV 百度风投、光源 L2F 创业者基金跟投）",
     "src": "https://news.ifeng.com/c/8s32SS1qOrn"
    }
   ],
   "key_data": [
    {
     "k": "GitHub stars",
     "v": "9,861",
     "date": "2026-11-03",
     "src": "https://github.com/Tiiny-AI/PowerInfer"
    },
    {
     "k": "Pocket Lab 众筹",
     "v": "Kickstarter 2026-03 上线 5 小时破 $100 万（BlockBeats）",
     "date": "2026-03",
     "src": "https://www.theblockbeats.info/flash/335829"
    }
   ],
   "note": "仓库已从 SJTU-IPADS 迁至 Tiiny-AI 组织。国内主体「本智激活」，孵化自上交 IPADS（虎嗅经凤凰网转载）。总部城市未核。README 称 Pocket Lab 跑 GPT-OSS-120B(int4) 约 20 tok/s。"
  },
  {
   "name": "PyTorch ExecuTorch",
   "layer": "s2",
   "url": "https://pytorch.org/executorch",
   "hq": "门洛帕克",
   "region": "海外",
   "type": "大厂产品线",
   "founded": 2022,
   "desc": "PyTorch 官方端侧运行时，覆盖手机到 MCU",
   "modality": "通用",
   "products": [
    "ExecuTorch"
   ],
   "devices": [
    "iPhone",
    "Android",
    "Mac",
    "Linux",
    "MCU",
    "树莓派"
   ],
   "chips": [
    "Arm CPU/Ethos-U",
    "Qualcomm Hexagon (QNN)",
    "Apple ANE (Core ML)",
    "MediaTek APU",
    "NXP NPU",
    "Samsung Exynos",
    "Intel NPU",
    "Vulkan GPU"
   ],
   "formats": [
    "ExecuTorch (.pte)"
   ],
   "license": {
    "v": "Meta BSD 许可（GitHub 标 NOASSERTION，LICENSE 文件为 BSD）",
    "kind": "开源",
    "src": "https://github.com/pytorch/executorch/blob/main/LICENSE"
   },
   "benchmarks": [
    {
     "k": "Llama 3.2 1B（4 位量化）· 三星 S24+ prefill",
     "v": "超过 350 tok/s（XNNPACK + KleidiAI，Arm CPU）",
     "date": "2024-10",
     "src": "https://pytorch.org/blog/unleashing-ai-mobile/"
    }
   ],
   "key_data": [
    {
     "k": "GitHub stars",
     "v": "5,152",
     "date": "2026-11-03",
     "src": "https://github.com/pytorch/executorch"
    }
   ],
   "product_line_of": "Meta",
   "public": true,
   "note": "LICENSE 版权行列有 Meta、Arm、Qualcomm、Apple、MediaTek、NXP、Samsung、Intel，可作后端贡献方线索。founded 取首次公开年份（凭背景知识，未核）。"
  },
  {
   "name": "腾讯 ncnn",
   "layer": "s2",
   "url": "https://github.com/Tencent/ncnn",
   "hq": "深圳",
   "region": "中国",
   "type": "大厂产品线",
   "founded": 2017,
   "desc": "手机端神经网络推理框架，无第三方依赖",
   "modality": "通用",
   "products": [
    "ncnn"
   ],
   "devices": [
    "Android",
    "iPhone",
    "Linux",
    "Windows",
    "Mac",
    "树莓派",
    "浏览器"
   ],
   "chips": [
    "Arm CPU",
    "Vulkan GPU",
    "Apple GPU"
   ],
   "formats": [
    "ncnn (param/bin)",
    "ONNX（经转换）"
   ],
   "license": {
    "v": "BSD-3-Clause（源码；GitHub 标 NOASSERTION）",
    "kind": "开源",
    "src": "https://github.com/Tencent/ncnn/blob/master/LICENSE.txt"
   },
   "key_data": [
    {
     "k": "GitHub stars",
     "v": "24,023",
     "date": "2026-11-03",
     "src": "https://github.com/Tencent/ncnn"
    }
   ],
   "product_line_of": "腾讯",
   "public": true,
   "note": "首次开源年份凭背景知识，未核。以视觉模型为主，LLM 支持程度未查。无基准。"
  },
  {
   "name": "阿里 MNN",
   "layer": "s2",
   "url": "https://github.com/alibaba/MNN",
   "hq": "杭州",
   "region": "中国",
   "type": "大厂产品线",
   "founded": 2019,
   "desc": "阿里端侧推理引擎，含 MNN-LLM 手机大模型",
   "modality": "通用",
   "products": [
    "MNN",
    "MNN-LLM",
    "MNN Chat"
   ],
   "devices": [
    "Android",
    "iPhone",
    "Linux",
    "Windows",
    "Mac"
   ],
   "chips": [
    "Arm CPU (KleidiAI)",
    "Qualcomm Adreno",
    "Apple GPU (Metal)",
    "MediaTek APU"
   ],
   "formats": [
    "MNN",
    "ONNX"
   ],
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://github.com/alibaba/MNN/blob/master/LICENSE.txt"
   },
   "key_data": [
    {
     "k": "GitHub stars",
     "v": "16,347",
     "date": "2026-11-03",
     "src": "https://github.com/alibaba/MNN"
    }
   ],
   "product_line_of": "阿里巴巴",
   "public": true,
   "note": "首次开源年份 2019 凭背景知识，未核。KleidiAI 已集成 MNN（Arm 开发者页）。无基准。"
  },
  {
   "name": "AMD Ryzen AI Software / Lemonade",
   "layer": "s3",
   "url": "https://ryzenai.docs.amd.com",
   "hq": "圣克拉拉",
   "region": "海外",
   "type": "大厂产品线",
   "founded": 2023,
   "desc": "Ryzen AI NPU 软件栈与 Lemonade",
   "modality": "通用",
   "products": [
    "Ryzen AI Software",
    "Lemonade",
    "ryzenai-server"
   ],
   "devices": [
    "Windows",
    "Linux"
   ],
   "chips": [
    "AMD XDNA NPU",
    "AMD Radeon iGPU",
    "AMD CPU"
   ],
   "formats": [
    "ONNX (OGA)",
    "GGUF (Lemonade)"
   ],
   "license": {
    "v": "RyzenAI-SW 为 MIT；Lemonade 为 Apache-2.0",
    "kind": "开源",
    "src": "https://github.com/amd/RyzenAI-SW/blob/main/LICENSE"
   },
   "key_data": [
    {
     "k": "GitHub stars（Lemonade）",
     "v": "6,096",
     "date": "2026-11-03",
     "src": "https://github.com/lemonade-sdk/lemonade"
    },
    {
     "k": "GitHub stars（RyzenAI-SW）",
     "v": "893",
     "date": "2026-11-03",
     "src": "https://github.com/amd/RyzenAI-SW"
    }
   ],
   "product_line_of": "AMD",
   "public": true,
   "note": "Ryzen AI Software 基于 ONNX Runtime GenAI，支持混合（NPU+iGPU）与纯 NPU 两种模式（lemonade-sdk/ryzenai-server 文档 https://github.com/lemonade-sdk/ryzenai-server ）。founded 凭背景知识。无官方 tokens/s 基准入库。"
  },
  {
   "name": "Apple Core ML / MLX",
   "layer": "s3",
   "url": "https://developer.apple.com/machine-learning/",
   "hq": "库比蒂诺",
   "region": "海外",
   "type": "大厂产品线",
   "founded": 2017,
   "desc": "Apple 端侧推理栈：Core ML 与 MLX",
   "modality": "通用",
   "products": [
    "Core ML",
    "coremltools",
    "MLX"
   ],
   "devices": [
    "iPhone",
    "Mac"
   ],
   "chips": [
    "Apple ANE",
    "Apple GPU (Metal)"
   ],
   "formats": [
    "Core ML (.mlpackage)",
    "MLX",
    "safetensors"
   ],
   "license": {
    "v": "MLX 为 MIT；coremltools 为 BSD-3-Clause；Core ML 框架随系统分发",
    "kind": "开源",
    "src": "https://github.com/ml-explore/mlx/blob/main/LICENSE"
   },
   "benchmarks": [
    {
     "k": "M4 Max · Nemotron-3 Nano 30B-A3B 4 位 解码（mlx-lm 0.31.3，第三方）",
     "v": "159.7 tok/s",
     "date": "2026-09",
     "src": "https://github.com/john-rocky/apple-silicon-llm-bench/blob/bd5feaf9fb1a5539046bb30789223d628d39d622/results/hybrid/nemotron3-nano-apple-silicon-bench.md"
    }
   ],
   "key_data": [
    {
     "k": "GitHub stars（MLX）",
     "v": "29,043",
     "date": "2026-11-03",
     "src": "https://github.com/ml-explore/mlx"
    },
    {
     "k": "GitHub stars（coremltools）",
     "v": "5,442",
     "date": "2026-11-03",
     "src": "https://github.com/apple/coremltools"
    }
   ],
   "product_line_of": "Apple",
   "public": true,
   "note": "founded 取 Core ML 首次公开年份 2017（凭背景知识，未核）；MLX 2023-12 发布。初稿提到的 iOS 27「Core AI」只来自社区转换页，未核，不写。"
  },
  {
   "name": "Arm KleidiAI",
   "layer": "s3",
   "url": "https://developer.arm.com/ai/kleidi-libraries",
   "hq": "剑桥（英国）",
   "region": "海外",
   "type": "大厂产品线",
   "founded": 2024,
   "desc": "Arm CPU 上的低比特 AI 内核库，被各框架内置",
   "modality": "通用",
   "products": [
    "KleidiAI",
    "Kleidi 库"
   ],
   "devices": [
    "Android",
    "iPhone",
    "Mac",
    "Linux",
    "树莓派"
   ],
   "chips": [
    "Arm CPU (SDOT/I8MM/SME2)"
   ],
   "formats": [
    "INT4/INT8 内核（随框架）"
   ],
   "license": {
    "v": "Apache-2.0（SPDX 标注）",
    "kind": "开源",
    "src": "https://github.com/ARM-software/kleidiai"
   },
   "benchmarks": [
    {
     "k": "ExecuTorch + KleidiAI · Llama 3.2 1B 4 位 · 三星 S24+ prefill",
     "v": "350+ tok/s；Cortex-A v9 上比无 KleidiAI 快最高 20%（Arm/Meta 联合博客）",
     "date": "2024-10",
     "src": "https://pytorch.org/blog/unleashing-ai-mobile/"
    }
   ],
   "key_data": [
    {
     "k": "GitHub stars（只读镜像；主仓库在 gitlab.arm.com/kleidi/kleidiai）",
     "v": "203",
     "date": "2026-11-03",
     "src": "https://github.com/ARM-software/kleidiai"
    },
    {
     "k": "已集成框架",
     "v": "MediaPipe(XNNPACK)、llama.cpp、PyTorch、ONNX Runtime、MNN、ExecuTorch、LiteRT",
     "date": "2026",
     "src": "https://developer.arm.com/ai/kleidi-libraries"
    }
   ],
   "product_line_of": "Arm",
   "public": true,
   "note": "founded 取 2024 凭背景知识，未核。stars 取自镜像，意义有限。"
  },
  {
   "name": "Google LiteRT / MediaPipe / LiteRT-LM",
   "layer": "s3",
   "url": "https://ai.google.dev/edge",
   "hq": "山景城",
   "region": "海外",
   "type": "大厂产品线",
   "founded": 2017,
   "desc": "TFLite 继任者 LiteRT 及 LLM 运行时",
   "modality": "通用",
   "products": [
    "LiteRT",
    "LiteRT-LM",
    "MediaPipe"
   ],
   "devices": [
    "Android",
    "iPhone",
    "浏览器",
    "Linux",
    "Mac",
    "Windows",
    "树莓派"
   ],
   "chips": [
    "Qualcomm Hexagon",
    "MediaTek APU (NeuroPilot)",
    "Google Tensor",
    "Apple GPU (Metal)",
    "Arm CPU (KleidiAI)"
   ],
   "formats": [
    "LiteRT (.tflite)",
    "LiteRT-LM (.litertlm)"
   ],
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://github.com/google-ai-edge/LiteRT/blob/main/LICENSE"
   },
   "benchmarks": [
    {
     "k": "Dimensity 9500 NPU · Gemma 3n E2B",
     "v": "prefill 1600+ tok/s，解码 28 tok/s（Google 官方）",
     "date": "2025-12",
     "src": "https://developers.googleblog.com/mediatek-npu-and-litert-powering-the-next-generation-of-on-device-ai/"
    },
    {
     "k": "M4 Max · Gemma-4-E2B .litertlm（Metal）解码，第三方",
     "v": "131.9 tok/s",
     "date": "2026-07",
     "src": "https://github.com/leehack/llamadart/blob/a395ebdd5a8c3076fd8d1cd746559b59094d48cf/website/versioned_docs/version-0.8.13/guides/backend-benchmarks.md"
    }
   ],
   "key_data": [
    {
     "k": "GitHub stars（MediaPipe）",
     "v": "37,540",
     "date": "2026-11-03",
     "src": "https://github.com/google-ai-edge/mediapipe"
    },
    {
     "k": "GitHub stars（LiteRT-LM）",
     "v": "6,811",
     "date": "2026-11-03",
     "src": "https://github.com/google-ai-edge/LiteRT-LM"
    },
    {
     "k": "GitHub stars（LiteRT）",
     "v": "3,564",
     "date": "2026-11-03",
     "src": "https://github.com/google-ai-edge/LiteRT"
    }
   ],
   "product_line_of": "Google",
   "public": true,
   "note": "founded 取 TensorFlow Lite 首次公开年份 2017（凭背景知识，LiteRT 改名为 2024 年，未核）。"
  },
  {
   "name": "Intel OpenVINO",
   "layer": "s3",
   "url": "https://docs.openvino.ai",
   "hq": "圣克拉拉",
   "region": "海外",
   "type": "大厂产品线",
   "founded": 2018,
   "desc": "Intel CPU/GPU/NPU 推理工具包",
   "modality": "通用",
   "products": [
    "OpenVINO",
    "OpenVINO GenAI"
   ],
   "devices": [
    "Windows",
    "Linux",
    "Mac",
    "浏览器"
   ],
   "chips": [
    "Intel NPU",
    "Intel CPU",
    "Intel Arc/Xe GPU"
   ],
   "formats": [
    "OpenVINO IR",
    "ONNX",
    "GGUF",
    "PyTorch"
   ],
   "license": {
    "v": "Apache-2.0",
    "kind": "开源",
    "src": "https://github.com/openvinotoolkit/openvino/blob/master/LICENSE"
   },
   "key_data": [
    {
     "k": "GitHub stars",
     "v": "11,111",
     "date": "2026-11-03",
     "src": "https://github.com/openvinotoolkit/openvino"
    },
    {
     "k": "当前版本",
     "v": "2026.4（GenAI 新增 Gemma 4 / Qwen 3.5 MTP 投机解码，NPU 支持 Kokoro-82M）",
     "date": "2026-09",
     "src": "https://www.intel.com/content/www/us/en/developer/tools/openvino-toolkit/whats-new.html"
    }
   ],
   "product_line_of": "Intel",
   "public": true,
   "note": "首次公开年份 2018 凭背景知识，未核。无官方统一 tokens/s 基准，省略。"
  },
  {
   "name": "Qualcomm AI Hub",
   "layer": "s3",
   "url": "https://aihub.qualcomm.com",
   "hq": "圣迭戈",
   "region": "海外",
   "type": "大厂产品线",
   "founded": 2024,
   "desc": "骁龙 NPU 部署栈：模型库、QNN、GenieX 运行时",
   "modality": "通用",
   "products": [
    "AI Hub",
    "QNN / Qualcomm AI Engine Direct",
    "GenieX"
   ],
   "devices": [
    "Android",
    "Windows",
    "Linux",
    "车机",
    "MCU"
   ],
   "chips": [
    "Qualcomm Hexagon",
    "Qualcomm Adreno",
    "Qualcomm Oryon CPU"
   ],
   "formats": [
    "QNN",
    "ONNX",
    "LiteRT",
    "GGUF (GenieX)"
   ],
   "license": {
    "v": "ai-hub-models 与 GenieX 为 BSD-3-Clause；QNN SDK 闭源",
    "kind": "开源",
    "src": "https://github.com/qualcomm/ai-hub-models/blob/main/LICENSE"
   },
   "key_data": [
    {
     "k": "GitHub stars（ai-hub-models）",
     "v": "1,251",
     "date": "2026-11-03",
     "src": "https://github.com/qualcomm/ai-hub-models"
    },
    {
     "k": "预优化模型数",
     "v": "175+（ASUS 2026-09 发布稿转述 AI Hub）",
     "date": "2026-09",
     "src": "https://www.ghacks.net/2026/09/18/asus-launches-the-ascent-qn10-an-80-tops-ai-mini-pc-powered-by-snapdragon-x2-elite/"
    },
    {
     "k": "骁龙 X2 系列 NPU 算力",
     "v": "最高 80 TOPS（Qualcomm 口径）",
     "date": "2025-11",
     "src": "https://www.notebookcheck.net/Hexagon-NPU-6-in-the-Snapdragon-X2-Elite-Extreme-80-TOPS-performance-that-is-up-to-95-faster-than-Apple-M4-and-122-faster-than-Intel-Lunar-Lake.1166576.0.html"
    }
   ],
   "product_line_of": "Qualcomm",
   "public": true,
   "note": "2026-07 开发者预览的 GenieX 即原 Nexa 团队产品（见 Nexa AI 卡）。founded 取 AI Hub 首次公开年份 2024（凭背景知识，未核）。无官方统一 tokens/s 基准。"
  },
  {
   "name": "华为 CANN / MindSpore Lite",
   "layer": "s3",
   "url": "https://www.hiascend.com/software/cann",
   "hq": "深圳",
   "region": "中国",
   "type": "大厂产品线",
   "founded": 2018,
   "desc": "昇腾/麒麟 NPU 的软件栈，2025 年宣布开源",
   "modality": "通用",
   "products": [
    "CANN",
    "MindSpore Lite"
   ],
   "devices": [
    "Android",
    "HarmonyOS",
    "Linux",
    "MCU"
   ],
   "chips": [
    "华为昇腾 NPU",
    "华为麒麟 NPU"
   ],
   "formats": [
    "MindIR",
    "OM",
    "ONNX"
   ],
   "license": {
    "v": "CANN 2025-08-05 宣布全面开源开放（具体协议本轮未核）",
    "kind": "开源",
    "src": "https://wap.yesky.com/enterprise/407/306407.shtml"
   },
   "key_data": [
    {
     "k": "开源时间",
     "v": "2025-08-05 昇腾计算产业发展峰会宣布 CANN 全面开源开放",
     "date": "2025-08",
     "src": "https://wap.yesky.com/enterprise/407/306407.shtml"
    }
   ],
   "product_line_of": "华为",
   "note": "CANN 1.0 于 2018-09 发布（百度百科，低权威）。MindSpore Lite 协议、仓库 stars 本轮未核（gitee 为主仓库）。无基准。"
  },
  {
   "name": "联发科 MediaTek NeuroPilot",
   "layer": "s3",
   "url": "https://neuropilot.mediatek.com",
   "hq": "新竹",
   "region": "中国",
   "type": "大厂产品线",
   "founded": 2018,
   "desc": "天玑 APU 的开发栈，接入 LiteRT 跑 Gemma",
   "modality": "通用",
   "products": [
    "NeuroPilot",
    "LiteRT NeuroPilot Accelerator"
   ],
   "devices": [
    "Android",
    "车机",
    "Linux"
   ],
   "chips": [
    "MediaTek APU (Dimensity 7300–9500)"
   ],
   "formats": [
    "LiteRT",
    "NeuroPilot DLA"
   ],
   "license": {
    "v": "闭源 SDK（商业许可）",
    "kind": "闭源",
    "src": "https://neuropilot.mediatek.com"
   },
   "benchmarks": [
    {
     "k": "Dimensity 9500 NPU · Gemma 3n E2B（经 LiteRT）",
     "v": "prefill 1600+ tok/s，解码 28 tok/s（Google 官方）",
     "date": "2025-12",
     "src": "https://developers.googleblog.com/mediatek-npu-and-litert-powering-the-next-generation-of-on-device-ai/"
    }
   ],
   "key_data": [
    {
     "k": "LiteRT 官方支持的天玑型号",
     "v": "7300 / 8300 / 9000 / 9200 / 9300 / 9400 / 9500",
     "date": "2026",
     "src": "https://ai.google.dev/edge/litert/next/mediatek"
    }
   ],
   "product_line_of": "联发科 MediaTek",
   "public": true,
   "note": "region 按联发科总部（台湾）暂归「中国」，建站时按规则再定。founded 为 NeuroPilot 品牌首次公开年份，凭背景知识，未核。许可页为 SDK 门户，未见公开许可条款。"
  },
  {
   "name": "Arduino",
   "layer": "s4",
   "url": "https://www.arduino.cc",
   "hq": "意大利",
   "region": "海外",
   "type": "创业",
   "founded": 2005,
   "desc": "开源开发板与 App Lab，UNO Q 集成高通芯片",
   "modality": "通用",
   "products": [
    "Arduino UNO Q",
    "Arduino App Lab",
    "Arduino IDE"
   ],
   "devices": [
    "MCU",
    "Linux",
    "树莓派"
   ],
   "chips": [
    "Qualcomm Dragonwing QRB2210",
    "STM32",
    "Arm Cortex-M"
   ],
   "formats": [
    "TFLite Micro",
    "Edge Impulse 导出"
   ],
   "license": {
    "v": "开源硬件；App CLI 为 GPL-3.0",
    "kind": "开源",
    "src": "https://github.com/arduino/arduino-app-cli"
   },
   "total_raised": "累计融资约 $54M（Arduino 官方博客口径，B 轮合计）",
   "funding": [
    {
     "date": "2022-06",
     "round": "B 轮",
     "amount": "$32M",
     "lead": "Bosch Ventures（Renesas、Arm、Anzu Partners 参与；二手来源）",
     "src": "https://profiles.crustdata.com/company/arduino"
    },
    {
     "date": "2023-09",
     "round": "B 轮追加",
     "amount": "$22M",
     "lead": "CDP Venture Capital、Anzu Partners 联合领投",
     "src": "https://blog.arduino.cc/2023/09/06/what-will-we-do-with-an-additional-22m/"
    }
   ],
   "key_data": [
    {
     "k": "开发者数",
     "v": "3300 万+（DLA Piper / Reuters 口径）",
     "date": "2025-10",
     "src": "https://www.dlapiper.com/en/news/2025/11/dla-piper-advises-qualcomm-group-on-acquisition-of-open-source-hardware-and-software-business"
    }
   ],
   "owner": "2025-10-07 被 Qualcomm 收购（金额未披露），保留独立品牌",
   "note": "收购见 Reuters 2025-10-07 https://www.reuters.com/world/asia-pacific/qualcomm-buys-open-source-electronics-firm-arduino-2025-10-07/ 。总部城市未取得一手出处，只写国家。UNO Q 售价 EUR 39 / USD 44 起（GSMA Intelligence）。"
  },
  {
   "name": "Edge Impulse",
   "layer": "s4",
   "url": "https://edgeimpulse.com",
   "hq": "圣何塞",
   "region": "海外",
   "type": "创业",
   "founded": 2019,
   "desc": "TinyML 开发平台：采数据、训模型、一键部署",
   "modality": "通用",
   "products": [
    "Edge Impulse Studio",
    "EON Compiler"
   ],
   "devices": [
    "MCU",
    "树莓派",
    "Jetson",
    "Linux",
    "车机"
   ],
   "chips": [
    "Arm Cortex-M",
    "Qualcomm Dragonwing",
    "NVIDIA GPU",
    "Syntiant",
    "Renesas",
    "Nordic"
   ],
   "formats": [
    "TFLite",
    "ONNX",
    "C++ 库"
   ],
   "license": {
    "v": "平台商业版；CLI 开源 BSD-3-Clause-Clear",
    "kind": "闭源",
    "src": "https://github.com/edgeimpulse/edge-impulse-cli/blob/master/LICENSE"
   },
   "valuation": {
    "v": "投前约 $2 亿",
    "date": "2021-12",
    "kind": "数据库口径（Crunchbase），公司未公布",
    "src": "https://www.crunchbase.com/funding_round/edge-impulse-series-b--344d9306"
   },
   "total_raised": "累计融资约 $54M（ZDNET 2021-12 报道口径，B 轮时）",
   "funding": [
    {
     "date": "2021-12",
     "round": "B 轮",
     "amount": "$34M",
     "lead": "Coatue（Canaan、Acrew、Fika、Momenta、Knollwood 跟投）",
     "src": "https://www.zdnet.com/article/edge-impulse-wants-to-bring-machine-learning-at-the-edge-to-everyone-announces-34m-series-b-funding/"
    }
   ],
   "key_data": [
    {
     "k": "开发者数",
     "v": "17 万+（Qualcomm 2025-03 公告口径）",
     "date": "2025-03",
     "src": "https://www.qualcomm.com/news/onq/2025/03/qualcomm-at-embedded-world-accelerating-digital-transformation-with-edge-ai"
    }
   ],
   "customers": [
    "NASA",
    "Oura",
    "Advantech",
    "Polycom"
   ],
   "owner": "2025-03-09 宣布与 Qualcomm 签署收购协议（金额未披露，交割以成交条件为准；交割日期本轮未查）",
   "note": "customers 为 2021 年 B 轮稿所列。总部城市凭背景知识，未逐项核。无基准。"
  }
 ]
}