Compress

Compress a model

Your compressed checkpoint, measured on your evaluation — same architecture, fewer parameters. We send a proposal.

Offering

A supervised compression run

We schedule and supervise the run: one compression level on your trained checkpoint, on high-performance compute we provision with our hardware partners, using methods from our own ongoing research.

•Compressed checkpoint — weights at the agreed density, same architecture as your original
•Plan file — the connectivity configuration and run parameters used for the compression
•Evaluation card and report — side-by-side results of original versus compressed on your named evaluation

Process

How it works

Plan or share your model

Use the planner on a Hugging Face model, or describe your checkpoint here.
Request a proposal

Tell us your data and evaluation. We review manually and send a proposal.
We run the compression

After you accept, we schedule the compute and supervise the run — with regular progress reports, not a single delivery at the end.
Receive your outputs

You get the compressed checkpoint, plan file, and evaluation report.

Evidence

Measured result: ViT-B/16

ViT-B/16 · ImageNet-1k · 34% uniform · one seed

Parameters (M)

Accuracy %

Dense baseline Compressed 84.51% → 77.58% · 86.6M → 30.5M (2.84×)

34% uniform · one seed. No latency or energy measurements reported. Other architectures and densities: measured on your evaluation.

Request a proposal

We review manually and send a proposal.