{
  "claim": "Edge0's README says the edge0-35b checkpoint is \"~23 GB\". It is not. Every tensor in Edge0/Edge0-35B-A3B-preview, read out of the four safetensors JSON headers by HTTP range request, totals 19,508,787,456 bytes — 19.51 GB, or 18.17 GiB. The paper says 19.5 GB and its own Figure 1 prints \"int4 experts (19.5 GB)\", so the README is the one number out of step with both the artifact and the paper. The same headers give the parameter count (34,660,610,688) and the memory floor: 1,389,394,176 bytes of non-expert weights that have to be resident no matter how aggressively the experts stream.",
  "method": "For each file: curl -sSL -X HEAD and read Content-Length after the CDN redirect (not the git-lfs pointer), cross-checked against the Hub API blob sizes. For each tensor: read the safetensors u64 header length from bytes 0-7 and the JSON header from bytes 8..n with an HTTP range request, then recover the unpacked width from the .scales group count (in = scales.shape[-1] x group_size) and verify it against the packed u32 width (in x bits / 32). No weights were downloaded — 1,757 tensors across four shards, resolved from about 1 MB of headers.",
  "source": "https://huggingface.co/Edge0/Edge0-35B-A3B-preview",
  "captured": "2026-09-18",
  "columns": [
    { "key": "component", "label": "component" },
    { "key": "params", "label": "parameters", "align": "right" },
    { "key": "bytes", "label": "bytes on disk", "align": "right" },
    { "key": "mib", "label": "MiB", "align": "right" },
    { "key": "resident", "label": "resident at decode?" }
  ],
  "rows": [
    { "component": "routed experts (40 layers x 256)", "params": 32212254720, "bytes": 18119393280, "mib": "17280.00", "resident": "no — streamed from SSD" },
    { "component": "attention (40 layers, 30 linear + 10 full)", "params": 1284188800, "bytes": 723784960, "mib": "690.26", "resident": "yes" },
    { "component": "lm_head (untied)", "params": 508559360, "bytes": 286064640, "mib": "272.81", "resident": "yes" },
    { "component": "embed_tokens", "params": 508559360, "bytes": 286064640, "mib": "272.81", "resident": "yes" },
    { "component": "shared expert (40 layers)", "params": 125911040, "bytes": 70865920, "mib": "67.58", "resident": "yes" },
    { "component": "router gates (40 layers, 8-bit)", "params": 20971520, "bytes": 22282240, "mib": "21.25", "resident": "yes" },
    { "component": "norms (bf16)", "params": 165888, "bytes": 331776, "mib": "0.32", "resident": "yes" },
    { "component": "= base checkpoint total", "params": 34660610688, "bytes": 19508787456, "mib": "18605.03", "resident": "—" },
    { "component": "prerouter heads (33 x 2,097,152, fp16)", "params": 69206016, "bytes": 138412032, "mib": "132.00", "resident": "yes" },
    { "component": "recovery LoRA (r=16, alpha=32, fp16)", "params": 21166080, "bytes": 42332160, "mib": "40.37", "resident": "yes" },
    { "component": "= non-expert resident floor", "params": 2538728064, "bytes": 1570138368, "mib": "1497.40", "resident": "yes" },
    { "component": "+ one K=4 decode step of experts (4 x 40 x 1,769,472 B)", "params": 503316480, "bytes": 283115520, "mib": "270.00", "resident": "transient" },
    { "component": "= floor with every layer's K=4 set held at once", "params": 3042044544, "bytes": 1853253888, "mib": "1767.40", "resident": "—" }
  ],
  "note": "Bytes are on-disk safetensors bytes, so a 4-bit affine group-64 tensor costs params/2 for the packed weights plus params/16 for the bf16 scales and biases (0.5625 bytes/param); the 8-bit router gates cost 1.0625. \"Parameters\" is the unpacked logical count. The vision tower that config.json declares (Qwen3_5MoeForConditionalGeneration, vision_config.depth 27, image_token_id, video_token_id) contributes zero of these bytes: all 1,757 tensors sit under the language_model. prefix and not one visual.* tensor ships. The config's mtp_num_hidden_layers: 1 likewise has no tensors behind it."
}