{
  "claim": "The shipped Hunyuan-A13B-Instruct checkpoint holds 80,393,183,232 parameters in 6,530 tensors. 96% of them sit in the 2,048 routed experts; there is no separate output-head tensor because the embedding is tied.",
  "method": "Read the 8-byte length prefix and the JSON header of each of the 33 safetensors shards with HTTP range requests (no weights downloaded), then summed the product of every tensor shape, grouped by tensor name.",
  "source": "huggingface.co/tencent/Hunyuan-A13B-Instruct, commit 290ddb9",
  "captured": "2026-09-26",
  "note": "Shapes are [out, in]. The router is the only float32 tensor: 32 layers x 64 experts x 4,096 = 8,388,608 parameters, which is the F32 slice Hugging Face lists for this repo.",
  "columns": [
    {
      "key": "family",
      "label": "tensor family"
    },
    {
      "key": "tensors",
      "label": "tensors",
      "align": "right"
    },
    {
      "key": "shape",
      "label": "shape",
      "mono": true
    },
    {
      "key": "dtype",
      "label": "dtype"
    },
    {
      "key": "params",
      "label": "parameters",
      "align": "right"
    }
  ],
  "rows": [
    {
      "family": "embed_tokens (tied with the output head)",
      "tensors": 1,
      "shape": "128167x4096",
      "dtype": "BF16",
      "params": 524972032
    },
    {
      "family": "attention q_proj",
      "tensors": 32,
      "shape": "4096x4096",
      "dtype": "BF16",
      "params": 536870912
    },
    {
      "family": "attention k_proj",
      "tensors": 32,
      "shape": "1024x4096",
      "dtype": "BF16",
      "params": 134217728
    },
    {
      "family": "attention v_proj",
      "tensors": 32,
      "shape": "1024x4096",
      "dtype": "BF16",
      "params": 134217728
    },
    {
      "family": "attention o_proj",
      "tensors": 32,
      "shape": "4096x4096",
      "dtype": "BF16",
      "params": 536870912
    },
    {
      "family": "QK-norm (query, key)",
      "tensors": 64,
      "shape": "128",
      "dtype": "BF16",
      "params": 8192
    },
    {
      "family": "router (mlp.gate.wg)",
      "tensors": 32,
      "shape": "64x4096",
      "dtype": "F32",
      "params": 8388608
    },
    {
      "family": "shared expert: gate, up, down",
      "tensors": 96,
      "shape": "3072x4096 / 4096x3072",
      "dtype": "BF16",
      "params": 1207959552
    },
    {
      "family": "routed experts: gate, up, down",
      "tensors": 6144,
      "shape": "3072x4096 / 4096x3072",
      "dtype": "BF16",
      "params": 77309411328
    },
    {
      "family": "RMSNorm (input, post-attention)",
      "tensors": 64,
      "shape": "4096",
      "dtype": "BF16",
      "params": 262144
    },
    {
      "family": "final RMSNorm",
      "tensors": 1,
      "shape": "4096",
      "dtype": "BF16",
      "params": 4096
    },
    {
      "family": "total",
      "tensors": 6530,
      "shape": "",
      "dtype": "",
      "params": 80393183232
    }
  ]
}