A guide from LMTimeline

A Timeline of Large Language Models (LLMs)

Large language models went from a research curiosity to the most consequential technology of the decade in roughly eight years. This is the throughline — the releases that actually moved the field — from the Transformer to today's reasoning models. For every release as it happens, see the live AI timeline.

The groundwork (2013–2016)

Before "LLM" was a household term, the pieces were assembling: word embeddings (word2vec, 2013) gave words mathematical meaning, and sequence-to-sequence models with attention made neural translation work. Attention — letting a model weigh every word against every other — was the idea that would soon be taken to its logical extreme.

2017 — The Transformer

Google's "Attention Is All You Need" introduced the Transformer, dropping recurrence entirely in favor of attention. It trained faster, scaled further, and became the foundation for essentially every large language model that followed.

2018–2019 — BERT and GPT-2

Google's BERT (2018) showed Transformers could deeply understand language, transforming search and NLP. OpenAI's GPT-2 (2019) took the generative path — a model fluent enough that its full weights were initially withheld over misuse fears, foreshadowing today's release debates.

2020 — GPT-3 and the scaling era

GPT-3's 175 billion parameters showed that scale alone produced new abilities — few-shot learning, coding, translation — without task-specific training. "Scaling laws" became the field's guiding principle, and the race to build bigger was on.

2022 — RLHF and ChatGPT

Reinforcement learning from human feedback (RLHF) taught models to follow instructions and stay helpful. Built on it, ChatGPT launched in November 2022 and became the fastest-growing consumer app ever — the moment LLMs went mainstream.

2023 — GPT-4, Claude, Gemini

The frontier became a contest. OpenAI's GPT-4, Anthropic's Claude and Google's Gemini pushed reasoning, reliability and multimodality, while context windows stretched from thousands of tokens toward millions.

2023–2024 — The open-weights wave

Meta's Llama models, along with Mistral, Qwen, DeepSeek and others, made capable LLMs freely downloadable — fueling a vast ecosystem of fine-tunes, on-device models and open research, and keeping pressure on the closed labs.

2025–today — Reasoning and agents

The newest LLMs think before they answer, spending compute on multi-step reasoning, and increasingly act as agents — calling tools, writing and running code, and orchestrating subagents on long tasks. They're also the first models whose releases are being shaped by governments: in 2026, frontier systems from OpenAI and Anthropic were staggered or limited to vetted partners over security concerns.

Keep reading

• The History of AI: A Timeline from 2006 to Today — the wider story around the models.
• The live AI timeline — filter to model releases, by maker or year, and subscribe for new entries.