Compression

The problem

Trained models are large, and the cost of storing and moving them grows with each generation. The usual way to make them smaller is to accept some loss of accuracy in return for space.

That trade has a cost of its own. A model that comes back slightly different is a different model, and the difference is hard to see until it matters.

What we set out to do

We asked whether trained models could take less storage and still come back exactly as they were. Not smaller with acceptable loss, but smaller with the contents exact.

The standard was strict from the start. If a single part came back different, the result did not count.

What we found

Trained models can take less storage with their contents exact.

Across several related families of model, storage fell while every part was restored exactly, and we recorded the gain for each model.

Along the way we established, in principle and in measurement, how far conventional exact compression of existing trained models can go. Those findings, and others from the same work, are set out on the Findings page.

What it is for

Exact compression can make trained models cheaper to store and move, without asking anyone to accept a model that is almost the same as the one they trusted.

CNA takes a different route: compression is part of the model from the outset, rather than applied to a finished one.

Where it stands

FCN Compression is the programme behind this work. Its papers were published in June 2026.

Conventional exact compression met the same limit in every model family we measured. Our own work takes models below their stored size with every part restored exactly, and the gain is recorded for each model.