Llama 2

Llama 2 is Meta AI's July 18, 2023 model family, the release that converted the accidental openness of the LLaMA leak into official strategy: weights freely downloadable under a community license permitting commercial use, with Microsoft as launch partner. It was the first openly licensed model family with a documented RLHF pipeline, making its paper the field's most complete public description of alignment training at the time.¹

Architecture and training

Llama 2 trained on two trillion tokens, doubled LLaMA's context to 4,096 tokens, and adopted grouped-query attention on the 70B model for serving efficiency. The chat variants layered SFT, iterative RLHF with two reward models (helpfulness and safety), and techniques like Ghost Attention for multi-turn consistency, all documented step by step in the paper.¹

License politics

The Llama Community License allows commercial use below a 700-million-user threshold and restricts using outputs to improve competitors: open weights, but not open source by OSI definition. The distinction ignited a definitional fight that still structures this wiki's open-weight coverage, with Mistral AI's Apache-2.0 releases and later fully open efforts from AI2 positioned explicitly against it.²³

Impact

Llama 2 became the default enterprise and research base of late 2023: the substrate of the fine-tune economy on Hugging Face, the baseline in every MMLU-era evaluation table, and the proof that a trillion-dollar company would arm the open ecosystem against the closed APIs of OpenAI and Anthropic. Its safety-tuned chat models also drew the era's canonical over-refusal complaints, a behavioral lesson successors tuned against.¹²

Lineage

Predecessor: LLaMA. Successors: Code Llama (August 2023), Llama 3 (2024). See Llama (model family).

References