I want the autonomy but local models of the size I would have the means to host wouldn't be capable enough. What usecases tend to suit these smaller models that tend to produce incorrect or otherwise flawed responses often? Could they work for anomaly detection and what would a rough architecture look like?