Flan-T5-XXL

Flan-T5-XXL is the largest variant of Google's Flan-T5 family, an 11B-parameter encoder-decoder model that applies instruction fine-tuning (FLAN) on top of the T5 v1.1 architecture. Released with the paper Scaling Instruction-Finetuned Language Models, it demonstrated that instruction tuning could give even encoder-decoder models strong zero-shot generalization.

It is released under the Apache 2.0 license and remains a popular choice for tasks like summarization, translation, and question answering where an encoder-decoder structure is a natural fit, as well as a backbone for vision-language systems such as BLIP-2.