Llama 3
Meta's open-weight LLM family, released April 2024. Llama 3 (8B + 70B), Llama 3.1 (July 2024, added 405B), Llama 3.2 (September 2024, multimodal + small).
Llama 3 405B was the first open-weight model in the GPT-4 class. License: allows commercial use under 700M MAU. Trained on 15T tokens. Llama 3.2 Vision (11B, 90B) added multimodality. Llama 4 (April 2025) introduced native MoE architecture. Hugely popular for fine-tuning + local inference. Mark Zuckerberg has pushed the open-weight strategy as a counterweight to OpenAI/Anthropic.