NullNet 19.98Q
19,980,000,000,000,000 logical 2-bit parameters.
NullNet is an intentionally useless packed-ternary Safetensors artifact.
Packing
The physical Safetensors dtype is U8. Each byte stores four 2-bit lanes:
00 -> 001 -> +110 -> -111 -> 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.
