{
  "claim": "The released checkpoint's own safetensors header settles what this article originally had to infer from the file's byte count. Every one of auk_base.safetensors' 420 tensors is F32 — the fp32 storage conclusion was right — and the header's shapes sum to 1,530,538,629 parameters. That is 13,644 below the file-size-divided-by-four estimate, because dividing the whole file by 4 also divides the 54,576-byte header by 4. AuK-Flash's checkpoint has the identical shape, tensor count and dtype.",
  "method": "HTTP range reads against huggingface.co: bytes 0-7 give the safetensors header length as a little-endian u64, bytes 8..8+n-1 give the JSON header. Parameter counts are the product of each tensor's declared shape, summed; no weight bytes were downloaded. The Qwen2.5-Omni-3B figure is the same read over its three shards, restricted to the tensors AuK keeps at runtime (the Thinker, after infer_auk.py does `del thinker.visual`).",
  "source": "https://huggingface.co/tencent/AuK",
  "captured": "2026-09-18",
  "columns": [
    { "key": "artifact", "label": "artifact" },
    { "key": "tensors", "label": "tensors", "align": "right" },
    { "key": "params", "label": "parameters", "align": "right" },
    { "key": "dtype", "label": "dtype" },
    { "key": "bytes", "label": "file bytes", "align": "right" }
  ],
  "rows": [
    { "artifact": "tencent/AuK — auk_base.safetensors", "tensors": 420, "params": 1530538629, "dtype": "F32 (all 420)", "bytes": 6122209092 },
    { "artifact": "  └ 10 × MMDiT dual-stream block", "tensors": 200, "params": 755284480, "dtype": "F32", "bytes": 3021137920 },
    { "artifact": "  └ 20 × DiT single-stream block", "tensors": 200, "params": 755284480, "dtype": "F32", "bytes": 3021137920 },
    { "artifact": "  └ peripheral (embeds, norms, proj_out, fusion)", "tensors": 20, "params": 19969669, "dtype": "F32", "bytes": 79878676 },
    { "artifact": "tencent/AuK — vae.safetensors", "tensors": 1137, "params": 159299797, "dtype": "F32 (all 1137)", "bytes": 637322604 },
    { "artifact": "tencent/AuK-Flash — auk_flash.safetensors", "tensors": 420, "params": 1530538629, "dtype": "F32 (all 420)", "bytes": 6122209092 },
    { "artifact": "Qwen2.5-Omni-3B Thinker, minus vision tower (resident)", "tensors": 924, "params": 4034780160, "dtype": "BF16 on disk", "bytes": 8069560320 },
    { "artifact": "  └ thinker.model (36-layer LM body)", "tensors": 434, "params": 3085938688, "dtype": "BF16", "bytes": 6171877376 },
    { "artifact": "  └ thinker.audio_tower", "tensors": 489, "params": 637676544, "dtype": "BF16", "bytes": 1275353088 },
    { "artifact": "  └ thinker.lm_head (allocated, unused)", "tensors": 1, "params": 311164928, "dtype": "BF16", "bytes": 622329856 },
    { "artifact": "  └ thinker.visual (deleted at load)", "tensors": 518, "params": 668684288, "dtype": "BF16", "bytes": 1337368576 }
  ],
  "note": "The block-level rows are the sums of the per-block tensor shapes, not a model of them: every one of the ten dual-stream blocks is exactly 75,528,448 parameters and every one of the twenty single-stream blocks exactly 37,764,224, which is what this article derived from config.yaml alone before the header was read. The `bytes` column for sub-rows is params × 4 (F32) or × 2 (BF16) and therefore excludes the header. Qwen2.5-Omni-3B's full repo is 5,537,120,672 parameters; AuK resides only the Thinker and deletes its vision tower, leaving 4,034,780,160."
}