Encoder-decoder
A design with two stacks: one reads the input, the other writes the output while looking back at it.
Also: sequence-to-sequence, seq2seq
The original 2017 transformer was an encoder-decoder built for translation, and earlier recurrent translation systems used the same split. The encoder reads the full source text in both directions, and the decoder generates the output token by token, using cross-attention to consult the encoder. Google's T5 and Meta's BART, both from 2019, are well-known encoder-decoder language models, and the design is still common in translation and speech recognition, including OpenAI's Whisper. For general chat, decoder-only models have largely taken over.