{
  "claim": "The browser build's first download is 1,844,177,822 bytes — 1.844 GB, of which 1,834,167,351 is weights. The Space's README says \"about 1.84 GB\" and that is exact, not rounded in its own favour. Two of the seven external-data shards are 267,386,880 bytes each, and 2 × 267,386,880 = 534,773,760 = 130,560 × 2,048 × 2: the whole FP16 embedding table, 510.00 MiB, sitting in four 127.5 MiB pieces exactly as the conversion notes describe.",
  "method": "curl -sSL -I against huggingface.co for every file in Mike0021/MiniCPM5-2B-ONNX at the revision the app pins (04a6c49fcba3a65a0351c92644c3a7e9d4343059), reading Content-Length after the CDN redirect rather than the git-lfs pointer. The sum of the eight onnx/ files reproduces the repo's usedStorage (1,834,167,351) to the byte. The source-checkpoint SHA-256 was read from the x-linked-etag header on openbmb/MiniCPM5-2B at revision abe115e887989b14f05e64a3b260648329324c3f.",
  "source": "https://huggingface.co/Mike0021/MiniCPM5-2B-ONNX",
  "captured": "2026-09-18",
  "columns": [
    { "key": "file", "label": "file" },
    { "key": "bytes", "label": "bytes", "align": "right" },
    { "key": "mib", "label": "MiB", "align": "right" },
    { "key": "holds", "label": "what it holds" }
  ],
  "rows": [
    { "file": "onnx/model_q4f16.onnx", "bytes": 318519, "mib": "0.30", "holds": "graph only — 480 nodes, no weights" },
    { "file": "onnx/model_q4f16.onnx_data", "bytes": 265703424, "mib": "253.39", "holds": "int4 transformer weights" },
    { "file": "onnx/model_q4f16.onnx_data_1", "bytes": 265420800, "mib": "253.12", "holds": "int4 transformer weights" },
    { "file": "onnx/model_q4f16.onnx_data_2", "bytes": 265420800, "mib": "253.12", "holds": "int4 transformer weights" },
    { "file": "onnx/model_q4f16.onnx_data_3", "bytes": 265420800, "mib": "253.12", "holds": "int4 transformer weights" },
    { "file": "onnx/model_q4f16.onnx_data_4", "bytes": 237109248, "mib": "226.12", "holds": "int4 transformer weights" },
    { "file": "onnx/model_q4f16.onnx_data_5", "bytes": 267386880, "mib": "255.00", "holds": "2 × 127.5 MiB FP16 embedding shards" },
    { "file": "onnx/model_q4f16.onnx_data_6", "bytes": 267386880, "mib": "255.00", "holds": "2 × 127.5 MiB FP16 embedding shards" },
    { "file": "tokenizer.json", "bytes": 9894271, "mib": "9.44", "holds": "byte-identical to the official repo's" },
    { "file": "tokenizer_config.json", "bytes": 102804, "mib": "0.10", "holds": "tool-history serialization corrected" },
    { "file": "chat_template.jinja", "bytes": 8154, "mib": "0.01", "holds": "Jinja-compatibility edits (official: 9,060 B)" },
    { "file": "config.json", "bytes": 822, "mib": "0.00", "holds": "adds transformers.js_config, 7 external shards" },
    { "file": "manifest.json", "bytes": 3656, "mib": "0.00", "holds": "source revision + SHA-256 of every file" },
    { "file": "generation_config.json", "bytes": 213, "mib": "0.00", "holds": "byte-identical to the official repo's" },
    { "file": "special_tokens_map.json", "bytes": 551, "mib": "0.00", "holds": "byte-identical to the official repo's" }
  ],
  "note": "For comparison, measured the same way: openbmb/MiniCPM5-2B-GGUF's Q4_K_M is 1,561,318,368 bytes and its F16 is 5,039,006,688; the official BF16 safetensors is 5,033,557,096 bytes, and its x-linked-etag SHA-256 (14fb8e7f… b64402c) is exactly the hash the conversion says it verified before starting. The FP16 embedding is 29.2% of the browser build's weight bytes; the same matrix is 150,405,120 bytes at Q4_K inside the GGUF."
}