“For domain-specific use cases, DeepSpec draft models should be fine-tuned to achieve better results, especially if the target model is expected to run in thinking mode.”