Transformer

The neural architecture behind modern LLMs: stacked self-attention and feed-forward layers processing all tokens in parallel. Introduced in 2017's 'Attention Is All You Need'.