{
  "claim": "Every one of the eleven file sizes on the Qwen3.8-35B-A3B GGUF card is exact. All eleven round-trip to the byte at three decimal places, the card states its unit convention out loud (1 GB = 1,000,000,000 bytes), and a vision projector now ships alongside the text quants. That is a direct fix of the failure this article opened with: the 2B release was sold on '~1 GB on a phone' against a smallest file of 1.312 GB.",
  "method": "curl -sSL -I against huggingface.co for every file in empero-ai/Qwen3.8-35B-A3B-Distill-GGUF, reading Content-Length after the CDN redirect rather than the git-lfs pointer. Parameter counts come from HTTP range reads of the 21 safetensors shard headers plus model-mtp.safetensors in empero-ai/Qwen3.8-35B-A3B-Distill — the 8-byte length prefix, then the JSON header — with each tensor's shape multiplied out; no weight bytes were downloaded.",
  "source": "https://huggingface.co/empero-ai/Qwen3.8-35B-A3B-Distill-GGUF",
  "captured": "2026-09-18",
  "columns": [
    { "key": "file", "label": "file" },
    { "key": "card", "label": "card says", "align": "right" },
    { "key": "bytes", "label": "measured bytes", "align": "right" },
    { "key": "check", "label": "bytes ÷ 1e9", "align": "right" }
  ],
  "rows": [
    { "file": "Qwen3.8-35B-A3B-IQ2_M.gguf", "card": "12.558 GB", "bytes": 12558245760, "check": "12.55824576" },
    { "file": "Qwen3.8-35B-A3B-Q2_K.gguf", "card": "13.839 GB", "bytes": 13838604160, "check": "13.83860416" },
    { "file": "Qwen3.8-35B-A3B-IQ3_M.gguf", "card": "16.340 GB", "bytes": 16339529600, "check": "16.33952960" },
    { "file": "Qwen3.8-35B-A3B-Q3_K_M.gguf", "card": "17.664 GB", "bytes": 17663774592, "check": "17.66377459" },
    { "file": "Qwen3.8-35B-A3B-IQ4_XS.gguf", "card": "19.628 GB", "bytes": 19627788160, "check": "19.62778816" },
    { "file": "Qwen3.8-35B-A3B-Q4_K_M.gguf", "card": "21.713 GB", "bytes": 21713462944, "check": "21.71346294" },
    { "file": "Qwen3.8-35B-A3B-Q5_K_M.gguf", "card": "25.348 GB", "bytes": 25347532448, "check": "25.34753245" },
    { "file": "Qwen3.8-35B-A3B-Q6_K.gguf", "card": "29.209 GB", "bytes": 29208731296, "check": "29.20873130" },
    { "file": "Qwen3.8-35B-A3B-Q8_0.gguf", "card": "37.802 GB", "bytes": 37802149536, "check": "37.80214954" },
    { "file": "Qwen3.8-35B-A3B-BF16.gguf", "card": "71.067 GB", "bytes": 71066994688, "check": "71.06699469" },
    { "file": "mmproj-Qwen3.8-35B-A3B-F16.gguf", "card": "0.899 GB", "bytes": 899283520, "check": "0.89928352" }
  ],
  "note": "Parameter accounting, measured the same way: 35,951,822,704 total across all 22 safetensors files, every tensor BF16. That splits into model.language_model 34,152,051,328 + lm_head 508,559,360 (= the card's '34.7B text path'), model.visual 446,571,248, and an mtp block of 844,640,768 in its own file. Active per token — dense body 1,813,885,568, shared experts 125,911,040, and 8/256 of 32,212,254,720 routed-expert parameters — is 2,946,429,568, so 'A3B' is honest. model.safetensors.index.json is not: its total_size (71,903,645,408) counts the MTP file's bytes while its total_parameters (35,107,181,936) excludes the MTP file's parameters, which is why the Hugging Face API prints a BF16 parameter count larger than the total."
}