Solenopsisbot/nullnet-19.98q

NullNet 19.98Q

Full version of Solenopsisbot/nullnet-mini-1.2q

19,980,000,000,000,000 logical 2-bit parameters.

Built by @Solenopsisbot, inspired by tsfrm/vacuum-16t.

NullNet is an intentionally useless packed-ternary Safetensors artifact.

Packing

The physical Safetensors dtype is U8. Each byte stores four 2-bit lanes:

• 00 -> 0
• 01 -> +1
• 10 -> -1
• 11 -> unused

Every stored lane is one of the first three values, so the payload genuinely represents ternary weights.

config.json declares a compressed-tensors 2-bit packed format. Hugging Face's Safetensors parameter counter therefore applies the physical-container packing factor: 8 / 2 = 4 logical parameters per U8 element.

Numbers

Logical parameters	19,980,000,000,000,000
Shards	10,000
U8 containers per shard	499,500,000,000
Logical weights per shard	1,998,000,000,000
Tensor payload per shard	499.500 GB
Logical tensor bytes across filenames	4.995 PB
Useful training	0 steps

All 10,000 shard files are byte-identical server-side copies of one Xet object.

The seed payload is periodic packed ternary data rather than an all-zero byte stream because an earlier enormous all-zero Xet reconstruction hit the CAS shard-registration limit. The weights are deterministic and untrained.

Why 19.98Q?

The current Hub parser supports at most 10,000 Safetensors shard filenames, and the Hub's documented hard per-file ceiling is 500 GB. This build leaves a 500 MB safety margin per file for the Safetensors header and service limits.

It is very nearly the honest 2-bit packed ceiling.

Capabilities

No.

Limitations

Approximately all of them.