{
  "claim": "kaggle-tpu-lab's README says of its 3-bit expert quantization: \"The output is that of a 3-bit-expert quantization, not the bf16 model... a formal comparison against the full model is on the list.\" This is the weight-space half of that comparison, which does not need a TPU. I pulled matched byte ranges of the same expert matrices out of Unsloth's UD-IQ3_XXS GGUF and Unsloth's BF16 GGUF, dequantized the first with the engine's own glm53.iqquant, and measured the error against the second. The format that carries the gate and up matrices in 41 of the 42 served layers, IQ2_S at 2.5625 bits per weight, reproduces its bf16 counterpart with a 25% relative Frobenius error and a cosine similarity of 0.970. The down matrices, at IQ3_S, land at 13.9%. The three layers Unsloth spent extra bits on (11, 12, 44) get IQ4_XS down projections at 7.6%.",
  "method": "For each tensor: parse the GGUF tensor index from the shard header by HTTP range read, compute the absolute byte offset of expert e row 0, then range-read 512 rows in the quantized file and the same 512 rows in the BF16 file. Dequantize the quantized rows with glm53.iqquant.dequant_rows (the engine's own reader, which I first checked against gguf.quants.dequantize by running the repo's own test_iqquant.py: 62 tests pass on CPU). rel_fro = ||W_q - W_bf16||_F / ||W_bf16||_F over the 512x n_in block; cos is the full-block cosine similarity; matvec is ||(W_q - W_bf16) X||_F / ||W_bf16 X||_F for X a 64-column standard normal draw at seed 0. Three experts per tensor (0, 100, 287). No full weight files downloaded: 34 MB of range reads in total.",
  "source": "https://huggingface.co/unsloth/GLM-5.3-Flash-GGUF",
  "captured": "2026-09-18",
  "note": "This is weight-space error, not end-to-end quality: a 25% relative error on a gate matrix does not mean the served model is 25% worse, and the routing, the attention stack and the shared expert are untouched by it. It is a lower bound on how far the served weights are from the ones every published GLM-5.3-Flash benchmark was run on, and it is the number the README's \"3-bit\" rounds off. bits/weight is the GGUF block size: 82 bytes per 256 weights for IQ2_S, 110 for IQ3_S, 136 for IQ4_XS.",
  "columns": [
    { "key": "tensor", "label": "tensor", "align": "left" },
    { "key": "expert", "label": "expert", "align": "right" },
    { "key": "qtype", "label": "format", "align": "left" },
    { "key": "bits", "label": "bits/w", "align": "right" },
    { "key": "rel_fro", "label": "rel Frobenius", "align": "right" },
    { "key": "cos", "label": "cosine", "align": "right" },
    { "key": "matvec", "label": "matvec rel", "align": "right" }
  ],
  "rows": [
    { "tensor": "blk.20.ffn_gate_exps", "expert": 0, "qtype": "IQ2_S", "bits": "2.5625", "rel_fro": "0.2500", "cos": "0.970034", "matvec": "0.2487" },
    { "tensor": "blk.20.ffn_gate_exps", "expert": 100, "qtype": "IQ2_S", "bits": "2.5625", "rel_fro": "0.2499", "cos": "0.970074", "matvec": "0.2503" },
    { "tensor": "blk.20.ffn_gate_exps", "expert": 287, "qtype": "IQ2_S", "bits": "2.5625", "rel_fro": "0.2501", "cos": "0.970011", "matvec": "0.2500" },
    { "tensor": "blk.20.ffn_down_exps", "expert": 0, "qtype": "IQ3_S", "bits": "3.4375", "rel_fro": "0.1385", "cos": "0.991102", "matvec": "0.1387" },
    { "tensor": "blk.20.ffn_down_exps", "expert": 100, "qtype": "IQ3_S", "bits": "3.4375", "rel_fro": "0.1384", "cos": "0.991111", "matvec": "0.1395" },
    { "tensor": "blk.20.ffn_down_exps", "expert": 287, "qtype": "IQ3_S", "bits": "3.4375", "rel_fro": "0.1397", "cos": "0.990931", "matvec": "0.1399" },
    { "tensor": "blk.11.ffn_gate_exps", "expert": 0, "qtype": "IQ3_S", "bits": "3.4375", "rel_fro": "0.1375", "cos": "0.991242", "matvec": "0.1364" },
    { "tensor": "blk.11.ffn_gate_exps", "expert": 100, "qtype": "IQ3_S", "bits": "3.4375", "rel_fro": "0.1375", "cos": "0.991239", "matvec": "0.1378" },
    { "tensor": "blk.11.ffn_gate_exps", "expert": 287, "qtype": "IQ3_S", "bits": "3.4375", "rel_fro": "0.1377", "cos": "0.991218", "matvec": "0.1394" },
    { "tensor": "blk.11.ffn_down_exps", "expert": 0, "qtype": "IQ4_XS", "bits": "4.2500", "rel_fro": "0.0764", "cos": "0.997078", "matvec": "0.0763" },
    { "tensor": "blk.11.ffn_down_exps", "expert": 100, "qtype": "IQ4_XS", "bits": "4.2500", "rel_fro": "0.0767", "cos": "0.997051", "matvec": "0.0762" },
    { "tensor": "blk.11.ffn_down_exps", "expert": 287, "qtype": "IQ4_XS", "bits": "4.2500", "rel_fro": "0.0763", "cos": "0.997086", "matvec": "0.0760" }
  ]
}