{
  "collectedAt": "2026-09-22",
  "supplement": true,
  "baseSnapshot": "scores-deepswe-1.1-2026-09-12.json",
  "supplementNote": "补充档：只向 deepswe_1_1 当前官方快照（scores-deepswe-1.1-2026-09-12.json）追加厂商自报行，不替换官方快照；官方榜出现对应行后本档应由官方行取代。官方榜 deepswe.datacurve.ai 截至 2026-09-22 无 grok-4.7 行、无任何 MiMo 行，因此这些不是覆盖而是补空缺。harness 归属是交叉验证推断、不是厂商明示：两家自报表里与官方榜重叠的对比机型读数都落在官方 mini-swe-agent 行的取整距离内——xAI 表 Grok 4.6 High 65.2 对官方 mini-swe-agent - grok-4-6 (high) 65.188（差 0.012）；小米卡片 Claude Opus 5 74.0 / GPT-5.6 Sol 73.0 / Claude Fable 5 70.0 对官方各自最高档 73.649 / 72.667 / 69.912（差 0.35 / 0.33 / 0.09）——据此判定两家 DeepSWE v1.1 列与官方榜同 harness 同口径。同一交叉验证在 Terminal-Bench 4.0 上不成立（小米卡片 Opus 5 49.0 对官方 51.82、GPT-5.6 Sol 39.9 对 37.27、Fable 5 42.4 对 44.55，偏差 2~3 分且正负不一致），说明小米的 TB4 是自跑且 harness 未署名，故其 TB4 自报（Pro 34.9 / Flash 28.8）不进本档、仅留存于 scores-mimo-v26-grok47-round1-2026-09-22.json 的 vendorBenchmarksNotBoards。",
  "boards": [
    {
      "boardId": "deepswe_1_1",
      "name": "DeepSWE v1.1",
      "metric": "Pass@1 %",
      "url": "https://deepswe.datacurve.ai/",
      "snapshotDate": "2026-09-03",
      "note": "与 scores-deepswe-1.1-2026-09-12.json 同一板面元数据；本档不改快照日期。"
    }
  ],
  "scores": [
    {
      "boardId": "deepswe_1_1",
      "model": "mimo-v2.6-pro",
      "variantLabel": "mini-swe-agent - mimo-v2-6-pro",
      "score": 71.9,
      "secondary": {
        "selfReported": true,
        "reportedBy": "Xiaomi",
        "agentHarness": "mini-swe-agent",
        "agentHarnessAttribution": "inferred-by-comparator-crosscheck",
        "reasoningEffort": null,
        "effortNote": "MiMo 系列无 reasoning_effort 参数，模型卡未设 effort 档；模型卡推荐采样 temperature=1.0 / top_p=0.95",
        "temperature": 1.0,
        "topP": 0.95,
        "contextWindowTokens": 1000000
      },
      "source": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL",
      "checkedAt": "2026-09-22",
      "note": "官方模型卡（MIT）第 3 节 Evaluation Results 表 DeepSWE v1.1 行：Pro 71.9 / Flash 67.9 / V2.5 Pro 19.0。同一模型的 DeepSWE 数字有三个互相矛盾的口径，按只记录不择一的规矩全部登记：①模型卡 71.9（本档采用，与 DeepSeek V4.1 Flash 自报同为模型卡口径）；②官方发布页 mimo.mi.com/docs/en-US/news/latest/v2-6 记 RL 曲线 58.4→72.6；③第三方 orcarouter 读小米 Live 面板终点 72.57（面板图注 mini-swe-agent, avg@3）。差异来自 checkpoint 选取、取整与 avg@k，②③未采用。官方榜无 MiMo 行。"
    },
    {
      "boardId": "deepswe_1_1",
      "model": "mimo-v2.6-flash",
      "variantLabel": "mini-swe-agent - mimo-v2-6-flash",
      "score": 67.9,
      "secondary": {
        "selfReported": true,
        "reportedBy": "Xiaomi",
        "agentHarness": "mini-swe-agent",
        "agentHarnessAttribution": "inferred-by-comparator-crosscheck",
        "reasoningEffort": null,
        "effortNote": "MiMo 系列无 reasoning_effort 参数，模型卡未设 effort 档；模型卡推荐采样 temperature=1.0 / top_p=0.95",
        "temperature": 1.0,
        "topP": 0.95,
        "contextWindowTokens": 1000000
      },
      "source": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL",
      "checkedAt": "2026-09-22",
      "note": "官方模型卡同表 Flash 行 67.9。同型矛盾口径：官方发布页记 RL 曲线 48.8→65.7、第三方读面板终点 65.68，均未采用（理由见 Pro 条）。这是 mimo-v2.6-flash 在本项目七张榜上的第一个分数——此前它 11 个采用点因无分数不进任何能力图。官方榜无 MiMo 行。"
    },
    {
      "boardId": "deepswe_1_1",
      "model": "mimo-v2.5-pro",
      "variantLabel": "mini-swe-agent - mimo-v2-5-pro",
      "score": 19.0,
      "secondary": {
        "selfReported": true,
        "reportedBy": "Xiaomi",
        "agentHarness": "mini-swe-agent",
        "agentHarnessAttribution": "inferred-by-comparator-crosscheck",
        "reasoningEffort": null,
        "effortNote": "MiMo 系列无 reasoning_effort 参数，模型卡未设 effort 档",
        "temperature": 1.0,
        "topP": 0.95
      },
      "source": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL",
      "checkedAt": "2026-09-22",
      "note": "同表 MiMo-V2.5 Pro 列 19.0。注意这是厂商在展示新版提升时给出的上一代基准值，方向性偏低的动机存在，读者应按厂商自报折价；与该模型已有的官方榜读数（AA 智力 25.9869、Code Arena 1475、Agent Arena -3.57）量级不矛盾。小米未给 mimo-v2.5（非 Pro）的 DeepSWE 数字，故不补。"
    },
    {
      "boardId": "deepswe_1_1",
      "model": "grok-4.7",
      "variantLabel": "mini-swe-agent - grok-4-7 (high)",
      "score": 71.0,
      "secondary": {
        "selfReported": true,
        "reportedBy": "xAI",
        "agentHarness": "mini-swe-agent",
        "agentHarnessAttribution": "inferred-by-comparator-crosscheck",
        "reasoningEffort": "high",
        "effortNote": "发布页基准表整列为 xHigh，DeepSWE 行带星号单独注明 high effort，故本行 effort 为 high 而非 xhigh"
      },
      "source": "https://x.ai/news/grok-4-7",
      "checkedAt": "2026-09-22",
      "note": "xAI 发布页（2026-09-21）基准表 DeepSWE v1.1 行 71.0%*，星注 high effort。同表 Grok 4.6 High 65.2 与官方榜 mini-swe-agent - grok-4-6 (high) 65.188 差 0.012，据此判定同 harness 同口径。官方榜无 grok-4.7 行，该模型此前在 DeepSWE 榜缺位。同表 Terminal-Bench 4.0 自报 38.0% 不采用——该榜已有官方行 Grok Build - Grok 4.7 (xhigh) 37.6%，官方读数优先于厂商自报；同表 Grok 4.6 TB4 20.3 与官方榜 Grok Build - Grok 4.6 (high) 20.3 完全一致，进一步佐证两表同源。"
    }
  ]
}