Technology
How Minutus compresses a model without retraining it.
Post-training compression fitted to the model's behavior and memory target.
Anchored to what the model actually does
Each layer is replaced by a compact approximation fitted to reproduce the original model's outputs, not just its weights. The fit accounts for the small input shifts that compressing earlier layers introduces, so errors do not pile up through the network.
Errors placed where they matter least
Not every weight error matters equally. We measure how sensitive the model's own predictions are to each part of the computation and spend the bit budget where it changes the output most. This is what lets 3-bit models keep their instruction-following ability.
Three building blocks
01
Low-bit backbone
Weights stored at 2-3 bits, using high-dimensional vector quantization in the 2-bit regime.
02
Low-rank correction
A small, precise correction term recovers what the low-bit backbone loses.
03
Embedding compression
The vocabulary tables, up to a third of a small model's memory, are compressed too, without measurable quality loss.
No retraining required
Compression uses a small sample of text and does not require a training dataset.
From research to production
Pilots deliver compressed models as Hugging Face checkpoints, evaluated on the customer's tasks. An optimized inference runtime for compressed formats is in development; today we report quality at a given storage size.