Llama
Meta's open-weights LLM family — Llama 1 (Feb 2023, leaked on 4chan), Llama 2 (July 2023, commercial OK), Llama 3 (April 2024), Llama 3.1 405B, Llama 4 (April 2025, MoE).
Decoder-only transformer with RMSNorm, SwiGLU, RoPE, GQA — the recipe that became the standard. Trained on ~15 T tokens. Open weights but the Llama licence restricts use beyond 700 M MAU.
Kicked off the "open weights" era — made local fine-tuning, RAG and agents possible without paying OpenAI. Competitors in open weights: Mistral, Qwen, DeepSeek, Gemma. Inference via vLLM, llama.cpp, Ollama.