{
  "claim": "Cactus's page gives Needle 3 a range of 8 MB at 4 layers to 29 MB at 20. The 20-layer file it actually publishes is 35,335,380 bytes \u2014 35.34 MB, 33.7 MiB \u2014 about 22% over the top of that range. The mixed-precision scheme in config.json (embedding and mHC at 4 bits, everything else at 2) explains why; the page's range reads as an all-2-bit estimate. Only the 20-layer file is published, so the smaller slices are not checkable the same way and I make no claim about them.",
  "method": "curl -sSL -I <resolve url> \u2014 Content-Length of the final response after the CDN redirect, not the git-lfs pointer stub.",
  "source": "https://huggingface.co/Cactus-Compute/needle3",
  "captured": "2026-09-18",
  "columns": [
    {
      "key": "artifact",
      "label": "artifact"
    },
    {
      "key": "bytes",
      "label": "bytes",
      "align": "right"
    },
    {
      "key": "mb",
      "label": "MB (10^6)",
      "align": "right"
    },
    {
      "key": "mib",
      "label": "MiB (2^20)",
      "align": "right"
    },
    {
      "key": "note",
      "label": "note"
    }
  ],
  "rows": [
    {
      "artifact": "needle3.cact (shipped, 20L)",
      "bytes": 35335380,
      "mb": 35.34,
      "mib": 33.7,
      "note": "the deployable container the size claim is about"
    },
    {
      "artifact": "page's stated ceiling (20L)",
      "bytes": 29000000,
      "mb": 29.0,
      "mib": 27.66,
      "note": "cactuscompute.com/needle, \"8 MB at 4 layers to 29 MB at 20\""
    },
    {
      "artifact": "checkpoints/needle3.safetensors",
      "bytes": 242047978,
      "mb": 242.05,
      "mib": 230.83,
      "note": "fp16 checkpoint you fine-tune from; not the deployed artifact"
    },
    {
      "artifact": "config.json",
      "bytes": 1273,
      "mb": 0.0,
      "mib": 0.0,
      "note": "declares scheme embedding=4,mhc=4,default=2"
    }
  ],
  "note": "The fp16 checkpoint reconciles exactly: an 8-byte length prefix plus a 8,248-byte JSON header leaves 242,039,722 bytes of tensor payload, and summing the 64 tensors in that header gives 242,039,722 bytes \u2014 the same number to the byte. Those tensors hold 121,019,861 parameters, all F16. config.json says total_parameters is 121,021,910, which is 2,049 more than the checkpoint contains (0.0017%). Too small to move any claim in the article \u2014 recorded because it was measured, not because it matters."
}