This isn’t talking about compaction. This refers to performance as the model is loaded with 500k to 1m tokens.