Choose the target
A student size for distillation, or a quantization level for shrinking.
ML · ships today
Compress a model
without losing it
Distill or quantize, then accept the smaller model only if measured quality holds.
Ships as compress/distill · compress/quantize
Make a model smaller or cheaper to serve, measure the quality difference on held-out evaluation, and let a person accept or reject the result.
The pipeline
Distill or quantize
Run the compression job on your GPUs inside the budget.
Measure the delta
Score the compressed model against the original on the same held-out set.
Accept or reject
A gate holds the result; smaller is only better if it still works.
What the record keeps
The source model, the target settings, the compressed artifact, the before-and-after scores, and the acceptance decision.
Lower serving cost
Smaller models mean fewer GPUs per request.
No silent regressions
The quality delta is measured, not assumed.
Compare candidates
compress/distill-compare runs several students side by side.
Ships as compress/distill, compress/distill-compare and compress/quantize. Supported model families depend on the installed operators.
See how failure can trigger a reviewed improvement cycle →Run it on your own infrastructure.
Apache 2.0 licensed, self-hosted. cairn pipeline run compress/quantize