GODEL is a large-scale pretrained dialog model that uses a Transformer encoder-decoder architecture. It is trained on 551M multi-turn dialogs from Reddit and 5M instruction and knowledge-grounded dialogs. The model supports efficient fine-tuning for new goal-directed dialog tasks and is accessible via Huggingface Transformers with a live demo available.