QATQ · Open-source compression for AI infrastructure teams
Move AI model state.
Carry fewer bytes.
Saving or moving an AI conversation means handling large files of model memory. QATQ shrinks that saved state for storage and transfer, then restores every byte exactly.
One measured transfer of model state
- Without compression
- 50.3 MB
- With QATQ
- 14.0 MB
72.2% fewer bytes transferred.
Why smaller is better: fewer bytes mean less storage space and less network bandwidth needed to move the same model state.
The original data was restored exactly. Results vary by model and workload.
The problem
Model memory is bulky to save and move.
AI models keep working data as a conversation grows. When you export that data to save a session or move it between servers, those bytes take up storage and network bandwidth.
The solution
Smaller files. The same data back.
QATQ compresses the exported data before you store or send it. After decoding, your system gets the exact original bytes back, with no loss from compression.
How it works
Save it. Shrink it. Restore it.
- 01
Export the model state
Your runtime saves its working memory, including the conversation data known as a KV cache.
- 02
Compress with QATQ
QATQ turns that exported data into a smaller file for storage or transfer.
- 03
Restore every byte
Decode the file and import the original data back into a compatible runtime.
QATQ works on exported files today. It requires a runtime that can export and re-import model state; it does not reduce GPU memory usage while the model is running.
The evidence
Fewer bytes in a tested server-to-server move.
In one documented cloud test, QATQ produced a smaller transfer than both LZ4 and Zstd, two common compression tools. The test also checked that the task could continue after the move.
Lower is better in the “Bytes transferred” column: it means less data sent over the network. A higher percentage in “Reduction” means more bytes saved compared with the uncompressed baseline.
| Compression | Bytes transferred | Reduction | Exact restore |
|---|---|---|---|
| Raw streamed blocks | 50,331,648 | baseline | Yes |
| LZ4 baseline | 28,739,217 | 43.0% smaller | Yes |
| Zstd baseline | 20,405,381 | 59.5% smaller | Yes |
| QATQ | 14,004,990 | 72.2% smaller | Yes |
Source: QATQ compression evidence, cloud live-migration test, 22 June 2026. These results apply to the tested setup; savings depend on the model, data and integration.