The spectrum of openness: fully proprietary (API-only, no weights, no details — GPT-4, Claude), open-weight (weights released, architecture described, but training data and code withheld — Llama, Mistral), and open-source (weights, code, data, and training recipe all public — rare, mostly academic). Most "open-source AI" is actually open-weight. The distinction matters for reproducibility, auditability, and legal liability.
Open models enable: transparency (you can inspect what the model does), privacy (your data never leaves your infrastructure), customization (fine-tune for your specific needs), cost control (no per-token fees), research (academia can study and improve models), competition (prevents monopoly), and reliability (no dependence on a provider's uptime or policy changes). The open-source community has demonstrated remarkable capability in building efficient inference (llama.cpp), fine-tuning tools (PEFT, TRL), and model variants.
Closed models enable: safety controls (the provider can enforce usage policies), responsible deployment (monitoring for misuse), rapid capability updates (users get improvements without redeployment), and accountability (a responsible entity behind the model). The safety argument is strongest at the frontier: the most capable models pose the most potential for misuse, and once weights are released, safety guardrails can be removed by anyone. This is why most frontier models remain API-only.