Declared size: 26,959,946,080,000,000 parameters. What is actually stored: one 500 GB shard, kept once and referenced by 6,740 identical paths (Xet deduplication). No weights exist; nothing here can be loaded or run.
The Giant
26,959,946,080,000,000 declared parameters.
Comparison
| Repository | Declared Parameters |
|---|---|
| tsfrm/vacuum-16t | 16.501T |
| VickM/leviathan-19t | 18.997T |
| Solenopsisbot/nullnet-19.98q | 19.98Q |
| yeths/The-Quettamind | 39.2Q |
| The Giant | 26.96Q |
How the declared size is produced
weights/master.safetensorsis one valid safetensors file: a header of 117U8tensors (114 × 2³², plus a tail) tiling[0, 499,999,000,000), followed by that many zero bytes (sparse on the build machine).- 6,739 more paths are server-side copies of it, so the bytes are stored once: 500 GB, not petabytes.
config.jsondeclarescompressed-tensors/pack-quantized,num_bits: 1→ 8 parameters per stored byte.6,740 × 499,999,000,000 × 8 = 26,959,946,080,000,000.
6,740 is not arbitrary: it is the largest shard count the Hub's summariser will still process for this repo. 6,745 shards freeze the counter.
Measured limits
| Limit | Value | How it was found |
|---|---|---|
| Max file size | 500.0 GB | 600 GB and 500 GB + header both → HTTP 422 |
| Max files per repo | 20,000 | "would contain 20006 files … over the limit of 20000" |
| Shards summarised | ≈ 6,740 here (parser allows 10,000) | 6,740 renders, 6,745 freezes |
| Params per stored byte | 8 (U8/I8 + num_bits: 1) |
F4 → 2, F6_E2M3 → 1.33, BOOL → 1 |
Fractional num_bits |
rejected | 0.5 / 0.25 / 0.1 → no metadata |
metadata.total_size |
ignored | an index claiming 10¹⁸ was recomputed from files |
metadata.total_parameters |
clamped down only | parser returns min(declared, computed) |
Overlapping ranges, oversized shape |
rejected | no metadata at all |
| Single tensor > 2³² elements | rejected | a 5×10¹¹-element tensor returns nothing |
| Duplicate index entries for one path | counted once | 200 aliases → one file counted |
The arithmetic ceiling of the method is 20,000 files × 500 GB × 8 = 80Q; the practical ceiling is the summariser's shard limit, which is what this repo sits against.
License
- Downloads last month
- 184
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support