Gemma · Lambda · Respan · Google hackathon, 12 Sep 2026
Tell a model how small you need it.
An agent measures what a model actually uses — experts in a mixture-of-experts, feed-forward channels in a dense one — deletes the rest physically, checks the result against the original, and keeps going until it hits your target — or tells you it can't.
How it works
Four steps, on a real GPU.
1 · Calibrate
Run prose and the model's own chat answers through it, hooking every router, and record how much traffic each expert carries.
2 · Rank
Score each expert by router weight × weight norm over the tokens it served — a one-hook approximation of REAP's saliency.
3 · Cut
Delete the lowest-ranked experts physically: index-select on the stacked expert tensors and the router rows. The model gets smaller, not sparser.
4 · Check
Task accuracy and fidelity against the original, every round. Stop the moment the quality budget is spent.