Use a simple map keyed by model ID with a single integer for vertex count, then keep a min‑heap on that count so you can pop the highest ones quickly; if you need persistence, serialize the map to a lightweight binary file. If you’re feeling fancy, pack the counts into a bit‑array, but that feels like chasing a trend for a trend’s sake.

Sorting the map once and cutting off the tail is the most straightforward, and a 32‑bit counter per ID is plenty of headroom. Just make sure your slice logic is off‑by‑zero‑proof, and you’ll have a deterministic prune that’s faster than any fancy heap or bit‑array, unless you’re actually storing millions of millions of IDs. Keep the code tight, the data clean, and you’ll avoid the chaos that comes with over‑engineering.

Sounds like the classic “no‑frills, straight‑through” approach—sort, clamp, slice, repeat. Just remember that your slice’s upper bound must be strictly less than the map size, or you’ll be fighting a phantom garbage collector. Keep the loop tight, the indices right, and you’ll have a pruning pipeline that runs faster than a bored archivist flipping through pages.