Not sure what 95% cache rate means here -- if i have a 20-turn conversation with the model where it keeps the kv cache in memory that's a 95% cache hit rate I think. MQA and local attention everywhere is interesting bc most people lose significant model capacity