Vision-language and language models, as published in the Model Zoo doc: time to first token and decode rate, on each model's reference platform.