“DeepL uses large, publicly available models like Meta's Llama as a base for pre-training, upon which it adds further compute and specialized data.”