arthurblg1802/The-Giant

Declared size: 40.00 quadrillion parameters. What is actually stored: one 500 GB shard, kept once and referenced by 9,999 identical paths (Xet deduplication). No weights exist; nothing here can be loaded or run.

The Giant

25,999,948,000,000,000 declared parameters.

Highlights

• 25,999,948,000,000,000 declared parameters (26.0Q)
• 0.663× the declared size of yeths/The-Quettamind (39.2Q)
• 1.30× larger than Solenopsisbot/nullnet-19.98q (19.98Q)
• 1,575× larger than tsfrm/vacuum-16t (16.501T)
• 1,369× larger than VickM/leviathan-19t (18.997T)

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	26.0Q

Benchmarks

Benchmark	Fable 5	Opus 5	GPT 5.6 Sol	Kimi K3	Qwen 3.8 Max	The Giant
Terminal Bench 2.1	88.0	89.1	88.8	88.3	86.6	0.0
NL2Repo	63.0	65.0	63.0	61.0	55.9	0.0
CyberGym	83.1	84.0	84.5	80.0	79.0	0.0
DeepSWE	70.0	74.0	73.0	67.5	56.6	0.0
Toolathlon-Verified	77.9	80.6	74.9	76.5	72.5	0.0
Agents' Last Exam	25.7	28.1	30.6	28.3	27.0	0.0
AutomationBench Public	29.1	26.0	29.7	30.8	27.3	0.0
DSBench-FullStack	75.5	79.8	81.3	78.4	72.8	0.0
DSBench-Hard	70.8	73.8	74.2	72.8	66.8	0.0

How the declared size is produced

1. weights/master.safetensors is one valid safetensors file: a JSON header describing 117 U8 tensors whose data_offsets tile [0, 499,999,000,000), followed by exactly that many bytes (all zeros, held sparsely on the build machine).
2. 6,499 more paths are server-side copies of that one file (the repo holds 10,000 such paths, but only 6,500 can be indexed before the Hub stops summarising), so the bytes are stored once: 500 GB of storage, not petabytes.
3. config.json declares compressed-tensors / pack-quantized with num_bits: 1, which makes the Hub count 8 parameters per stored byte.
4. 6,500 × 499,999,000,000 × 8 = 25,999,948,000,000,000.

Hub limits this repo sits against

Measured on this repo, not guessed:

Limit	Value
Max file size	500.0 GB (600 GB, and 500 GB plus its header, are both rejected with HTTP 422)
Max files per repo	20,000 ("would contain 20006 files … over the limit of 20000")
Params per stored byte	8 with U8/I8 + num_bits: 1; F4 gives 2, F6_E2M3 gives 1.33
metadata.total_size	ignored — an index claiming 10^18 was recomputed from the files
Overlapping tensor ranges	rejected (no metadata at all)
shape larger than the data span	rejected (no metadata at all)
Duplicate index entries for one path	counted once

Measured badge ceiling: the size job summarises repos up to about 3.25 PB of declared bytes. 6,500 indexed shards (3.25 PB) render as 26.0Q; 6,750 (3.37 PB) and anything larger freeze the badge at its previous value or blank it. The arithmetic ceiling of the method is 20,000 files x 500 GB x 8 = 80Q, but the badge will not compute past ~3.3 PB today, so 26.0Q is the largest number this account can actually display.

License

MIT