As a rule of thumb inference offered by the model labs are closer to the "true implementation" compared to third parties. They have other problems though.