Encoder-decoder
T5, BART and the original transformer: read with one stack, write with another. 19 models on EveryLLM so far.
Encoder-decoder models use two stacks: an encoder reads the input in full, and a decoder writes the output one token at a time while attending to the encoder's results. This was the shape of the original 2017 transformer, built for translation, and it suits tasks that map one text to another, such as translation, summarisation and speech-to-text. Google's T5, Facebook's BART and OpenAI's Whisper are the best-known examples.
T5 and text-to-text
Google's T5 (October 2019) framed every task as turning one string into another, so translation, summarising and even classification share one model, with the task named at the start of the input. Its largest version had 11 billion parameters, trained on C4, a cleaned web crawl of about 750 GB. Flan-T5 (2022) added instruction tuning, and the multilingual mT5 covered 101 languages.
Where they are used now
Decoder-only models took over general chat because they are simpler to scale, but encoder-decoders remain strong where the input is fixed and the output comparatively short. OpenAI's Whisper (September 2022), a speech recognition model, uses an encoder for audio and a decoder for text. Google's T5Gemma (2025) converted Gemma decoder models into encoder-decoders and reported better quality for the same inference cost on several tasks.