KATOYING 
KTY-R002 · small model density
KATOYING TERNING
How small can a model become before it stops preserving the language it was trained to use?
In simple words
Many numbers become three choices.
A normal model stores many precise numbers inside its weights. This experiment asks a smaller question: can every large matrix live mostly with only three values?
The useful result is not just the smaller file. The ternary version preserved more of the expected symbolic structure than its full-precision reference on this miniature. Pushing further to binary or half-bit destroyed that structure.
Measured cliff
Smaller is not always better.
| Version | Density | Structure match | Result |
|---|---|---|---|
| Full-precision master | 16.01 bpw | 39.10% | Reference |
| Ternary training | 1.609 bpw | 53.20% | PASS |
| Binary | 1.009 bpw | 22.86% | COLLAPSE |
| Block VQ | 0.510 bpw | 10.48% | COLLAPSE |
The boundary measured here is between 1.609 and 1.009 bits per weight. This result belongs to this small symbolic model; it does not automatically transfer to every model.
On a phone
The artifact runs on an S21 Ultra.
The ternary file is 2.84 times smaller than the same architecture in Q4. Its current ARM kernel is still 11 to 16% slower in generation. The next honest speed gain is a better runtime kernel, not blind compression.
Boundary
What this does not prove.
It proves
- A 2.65 MiB ternary artifact was exported and matched across a workstation and an S21 by SHA-256.
- All 30 large matrices use the ternary runtime format.
- The 1,050-sequence validation and phone measurements produced the numbers above.
It does not prove
- A general-purpose AI agent.
- Reliable human conversation or tool calling.
- That 1-bit or 0.5-bit compression is currently usable.
- That this private training method should be published.
Public references
Where the field meets the experiment.
- Microsoft BitNet reference implementationOfficial open implementation for efficient low-bit model inference.
- The Era of 1-bit LLMs: All Large Language Models are in 1.58 BitsPrimary BitNet b1.58 research paper.
- llama.cpp quantization documentationPublic reference for runtime quantization tooling and formats.