Llama 2, Deliberately Open

By Satwik · April 8, 2026

Meta released Llama 2 in July 2023, and unlike the original it was open by design rather than by leak. The models came with published weights, a permissive-ish community license allowing commercial use with some conditions, chat-tuned variants, and an unusually detailed paper covering the pretraining, the supervised fine-tuning, the reinforcement learning from human feedback, and the safety work. Meta shipped it in partnership with a major cloud provider and framed openness as a deliberate strategy.

The release matters as the establishment of a genuinely open, reasonably capable, commercially usable model family from a major lab. It legitimized building products and research on open weights and became the backbone for a huge amount of downstream fine-tuning, including many of the specialized models that followed.

For security the Llama 2 documentation is a valuable primary source because Meta was relatively forthcoming about safety methodology, including red-teaming and the tension between helpfulness and harmlessness. But the deliberate-openness posture makes the irreversibility problem explicit rather than accidental. Once safety-tuned open weights are published, anyone can fine-tune the guardrails away, and the community demonstrated exactly that within weeks by producing uncensored variants. So the safety tuning documented in the paper protects the default checkpoint, not the capability, which is fully in the wild and adaptable. Llama 2 is the clearest case study of the central open-weights tradeoff: broad access accelerates beneficial research, transparency, and competition, while conceding that behavioral safety measures are removable and that the underlying capability, once released, is permanent. It made that tradeoff a matter of stated policy rather than an after-the-fact accident.