One thing I noticed with local models is that conversation behavior tends to drift over time as well.

Even when running locally, the model often starts structured but gradually becomes more verbose or explanatory in longer threads.

Curious if others have seen similar behavior when using local setups.