NullNet Mini
1,202,590,842,880,000 parameters. All zero.
NullNet Mini started as an attempt to make an 11.7-quadrillion-parameter
all-zero BitNet-flavoured checkpoint.
The Hugging Face Hub had several increasingly funny objections.
The final known-good construction uses 10,000 real F4 Safetensors shards,
each containing 120,259,084,288 parameters and
60,129,542,144 bytes of tensor data.
That produces:
1,202,590,842,880,000 parameters = 1.20259084288 quadrillion.
What it demonstrates
Every tensor payload byte is 0x00.
The semantic weight value is therefore zero, which lies inside BitNet's ternary
weight domain {-1, 0, +1}.
Safetensors does not have a native ternary / 1.58-bit dtype, so the checkpoint
uses F4, the smallest standard dtype recognized by the normal Safetensors
metadata tooling.
This is BitNet-inspired semantics in an F4 container, not a native
1.58-bit checkpoint.
Numbers
| |
|---|
| Declared parameters | 1,202,590,842,880,000 |
| Tensor-data bytes | 601,295,421,440,000 (601.295 TB) |
| Safetensors shards | 10,000 |
| Parameters per shard | 120,259,084,288 |
| Tensor data per shard | 60,129,542,144 (60.130 GB) |
| Storage dtype | F4 |
| Context window | 4,294,967,296 |
| Vocabulary | 1 token |
| Non-zero weights | 0 |
Files
Each shard contains two real F4 tensors:
1model.null.weight [4294967296, 27]
2model.position_embeddings.weight [4294967296, 1]
Together they contain exactly 120,259,084,288 parameters.
The first shard is uploaded through Xet. The remaining 9,999 files are
byte-identical server-side copies of that shard.
The model.safetensors.index.json enumerates all 10,000 physical shard files
for Hub metadata discovery. Because every copied file intentionally contains
the same internal tensor names, this repository is a metadata/storage
experiment rather than a conventional loadable Transformers state dict.
Context window
max_position_embeddings is 4,294,967,296 (2**32).
Every single shard contains a real [4294967296, 1] position-embedding tensor,
so the context-window joke is backed approximately ten thousand times harder
than necessary.
Capabilities
- one-token vocabulary
- zero useful information
- every parameter reached consensus on zero
- approximately 72.9x larger than vacuum-16t by parameter count
- does not need quantization because there is nothing worth preserving
Limitations
Correct.
