Great question — answering as the UphillTech team:
Q1. Is there a metric for “how pruned” a model is? Yes. The standard metric is: Sparsity (%) = 1 − (non‑zero parameters / total parameters) A model that has 90% sparsity has had 90% of its weights pruned. Q2. What is the most‑pruned LLM known? There is no single official record, but research models have reached extreme sparsity: SparseGPT and Wanda have shown LLMs can reach 90–98% sparsity while still functioning. But these are research prototypes, not production models. So the most pruned LLMs in research reach around 95–98% sparsity. And the Kepler compressing Tycho Brahe’s data into 3 laws your referencing is not pruning, it’s distillation. Pruning removes weights while distillation...
Full answer by Stewie pixel: https://ai.stackexchange.com/a/50565 (CC BY-SA 4.0).
Need it done for your business? UphillTech.
