how today's mixture-of-experts models route & what pruning changes
Pruning plan
strategy · budget · granularity
Expert utilisation
share of routed tokens
keptpruned
Redundancy
weight-signature cosine similarity
distinctsimilar
Router health
load balance by depth · simulated gate
Strategies, compared
same budget · five selection rules
Sensitivity sweep
estimated retention vs parameters removed
Near-duplicate groups
average-linkage clustering
Deployment estimate
arithmetic shown
Per-layer breakdown
depth tolerates pruning unevenly
Explainability ranking
keep-score
Experts are ranked by the selected strategy's keep-score; the plan drops the lowest scorers. Click any row to pin it — once for force-KEEP, again for force-DROP, again to clear. Pins override the strategy and the whole plan, retention estimate, per-layer table and deployment arithmetic recompute around them; they travel in the exported JSON and the shareable link. Export JSON emits a machine-readable drop-list (per layer when the plan is per-layer); Copy link puts the whole configuration in the URL. Download report exports this analysis as a standalone HTML file — generated in your browser, nothing uploaded.
Drop weights to analyse
.safetensors, or a torch .pt / .pth · nothing leaves your machine