What exactly is REAP | LMStudio guide

TL;DR

• 1

  REAP is targeted compression for mixture-of-experts models: Instead of shrinking every weight equally, router-weighted activation pruning watches which experts light up on calibration prompts and removes the parts you do not need.

• 2

  The size savings are huge: Sharif says a 60 GB model can drop to roughly 12.5% of its original size with recommended pruning plus quantization, and a 1.5 TB GLM 5.1 setup was brought down to about 370 GB.

• 3

  Big models are worth cutting down because their remaining weights are often better than a native small model: His argument is that frontier models like GLM 5.1 benefit from better data, better engineers, and more training tokens, so preserving 30% of a huge model can beat starting with a much smaller one.

• 4

  Quantization and pruning are different tools: Quantization changes 16-bit weights into smaller representations like 4-bit, while pruning literally removes parts of the model, and both can be combined depending on your hardware budget.

• 5

  The economics of hosted AI look shaky: Sharif claims the same task can cost 5 to 11 times more than it did three years ago because of thinking tokens, tool calls, and larger prompts, and says heavy users on $20 to $200 subscriptions are often being subsidized.

• 6

  You do not necessarily need to retrain the router after pruning experts: In the Q&A, he says the router can repoint to the remaining salient experts out of the box, though retraining can improve performance.