{
  "protocol": "younis-ai-lab/model-memory-estimator",
  "version": "1.0.0",
  "calculation": {
    "parameterCountScale": 1000000000,
    "bytesPerGiB": 1073741824,
    "bitWidths": [
      16,
      8,
      4
    ],
    "roundingDecimalPlaces": 1,
    "formula": "parameterBillions * 1_000_000_000 * bitsPerWeight / 8 / 1_073_741_824"
  },
  "models": [
    {
      "input": {
        "name": "DeepSeek-R1",
        "parameterBillions": 671,
        "activeParameterBillions": 37
      },
      "estimates": [
        {
          "bitsPerWeight": 16,
          "idealWeightGiB": 1249.8
        },
        {
          "bitsPerWeight": 8,
          "idealWeightGiB": 624.9
        },
        {
          "bitsPerWeight": 4,
          "idealWeightGiB": 312.5
        }
      ]
    },
    {
      "input": {
        "name": "R1-Distill-Llama-70B",
        "parameterBillions": 70,
        "activeParameterBillions": null
      },
      "estimates": [
        {
          "bitsPerWeight": 16,
          "idealWeightGiB": 130.4
        },
        {
          "bitsPerWeight": 8,
          "idealWeightGiB": 65.2
        },
        {
          "bitsPerWeight": 4,
          "idealWeightGiB": 32.6
        }
      ]
    },
    {
      "input": {
        "name": "R1-Distill-Qwen-32B",
        "parameterBillions": 32,
        "activeParameterBillions": null
      },
      "estimates": [
        {
          "bitsPerWeight": 16,
          "idealWeightGiB": 59.6
        },
        {
          "bitsPerWeight": 8,
          "idealWeightGiB": 29.8
        },
        {
          "bitsPerWeight": 4,
          "idealWeightGiB": 14.9
        }
      ]
    },
    {
      "input": {
        "name": "R1-Distill-Qwen-14B",
        "parameterBillions": 14,
        "activeParameterBillions": null
      },
      "estimates": [
        {
          "bitsPerWeight": 16,
          "idealWeightGiB": 26.1
        },
        {
          "bitsPerWeight": 8,
          "idealWeightGiB": 13
        },
        {
          "bitsPerWeight": 4,
          "idealWeightGiB": 6.5
        }
      ]
    },
    {
      "input": {
        "name": "R1-Distill-Llama-8B",
        "parameterBillions": 8,
        "activeParameterBillions": null
      },
      "estimates": [
        {
          "bitsPerWeight": 16,
          "idealWeightGiB": 14.9
        },
        {
          "bitsPerWeight": 8,
          "idealWeightGiB": 7.5
        },
        {
          "bitsPerWeight": 4,
          "idealWeightGiB": 3.7
        }
      ]
    },
    {
      "input": {
        "name": "R1-Distill-Qwen-7B",
        "parameterBillions": 7,
        "activeParameterBillions": null
      },
      "estimates": [
        {
          "bitsPerWeight": 16,
          "idealWeightGiB": 13
        },
        {
          "bitsPerWeight": 8,
          "idealWeightGiB": 6.5
        },
        {
          "bitsPerWeight": 4,
          "idealWeightGiB": 3.3
        }
      ]
    },
    {
      "input": {
        "name": "R1-Distill-Qwen-1.5B",
        "parameterBillions": 1.5,
        "activeParameterBillions": null
      },
      "estimates": [
        {
          "bitsPerWeight": 16,
          "idealWeightGiB": 2.8
        },
        {
          "bitsPerWeight": 8,
          "idealWeightGiB": 1.4
        },
        {
          "bitsPerWeight": 4,
          "idealWeightGiB": 0.7
        }
      ]
    }
  ],
  "boundaries": {
    "scope": "只按型号中的十进制名义参数量估算理想连续权重存储，并以二进制 GiB 表示。",
    "checkpointIdentity": "名义参数量不等于具体 checkpoint 的实际张量数或文件字节数。",
    "moe": "MoE 激活参数量描述每 token 的计算稀疏度，不能替代完整权重存储所需的总参数量。",
    "excluded": [
      "量化 scale、zero-point、分组元数据与文件格式开销",
      "KV Cache、激活、运行时 workspace 与通信 buffer",
      "操作系统、推理服务、容错副本与故障恢复余量",
      "设备放置、并发、上下文长度、吞吐、延迟、功耗与任务正确性"
    ]
  }
}