{
  "claim": "Every published laya-coreml bundle freezes its option capacity at export time. K is never flexible in any of the six: it is 32 in five of them and 4 in the Snake bundle, whose four slots exist because Snake has four directions. The sequence length is flexible only in the three ordinary Core ML exports, and the Neural Engine bundles freeze the batch size and the sequence length too.",
  "method": "Read out of the repository's own release inventory, which records the exported shape dictionary written into each bundle's coreml_config.json at conversion time. B is the exported batch size (questions per forward pass), L the sequence capacity in tokens, K the max_options marker-slot count. 'length buckets' is the number of enumerated sequence shapes the export carries; a dash means the length is a single literal. Bundle size is the whole downloadable repository in decimal MB, including the tokenizer and, for the Neural Engine bundles, the host-side embedding and action-head tensors.",
  "source": "https://github.com/mizorewww/laya-coreml/blob/main/benchmarks/results/hub-release.json",
  "captured": "2026-09-20",
  "note": "Nothing here is quantised except the W8 bundle, which is weight-only K-means palettisation with FP16 compute. The 122 MB it saves against the FP16 ANE bundle is the body package shrinking from 251.91 to 129.29 MB; the other ~428 MB of both bundles is the host embedding table and tokenizer, which never reach the Neural Engine.",
  "columns": [
    { "key": "bundle", "label": "hugging face bundle", "align": "left" },
    { "key": "b", "label": "B", "align": "right", "mono": true },
    { "key": "l", "label": "L", "align": "right", "mono": true },
    { "key": "k", "label": "K", "align": "right", "mono": true },
    { "key": "buckets", "label": "length buckets", "align": "right", "mono": true },
    { "key": "mb", "label": "bundle MB", "align": "right", "mono": true }
  ],
  "rows": [
    { "bundle": "aac6fef/laya-coreml", "b": 1, "l": "16-512", "k": 32, "buckets": 9, "mb": 847.2 },
    { "bundle": "aac6fef/laya-multilingual-coreml", "b": 1, "l": "16-1024", "k": 32, "buckets": 11, "mb": 679.9 },
    { "bundle": "aac6fef/laya-typed-decisions-coreml", "b": 1, "l": "16-1024", "k": 32, "buckets": 11, "mb": 848.2 },
    { "bundle": "aac6fef/laya-multilingual-coreml-ane", "b": 1, "l": 96, "k": 32, "buckets": "—", "mb": 679.9 },
    { "bundle": "aac6fef/laya-multilingual-coreml-ane-w8", "b": 1, "l": 96, "k": 32, "buckets": "—", "mb": 557.3 },
    { "bundle": "aac6fef/laya-multilingual-coreml-snake", "b": 3, "l": 64, "k": 4, "buckets": "—", "mb": 678.5 }
  ]
}