A language model can be available for others to run and adapt without revealing how it was made. An open-weight model shares its learned parameters—the values produced during training—but may keep its training data, methods, or pipeline closed. Its license determines what others are allowed to do with the release. This differs from open source, which strictly means a license permits full use, modification, and redistribution. section 11.4
Shared weights can make experimentation and customization more accessible. After Meta released Llama 1 under a noncommercial research license, its weights leaked online; Stanford’s Alpaca then showed how they could be repurposed into a chatbot at modest cost. Other fine-tuned models soon followed. Some drew on instructions generated by commercial systems, blurring a simple division between “open” and “closed” progress. section 11.4
Availability brings trade-offs:
- A release may support research and reproducibility, but shared weights alone do not reveal every training choice or settle what its license permits. section 11.4 section 3.17
- People deploying an openly available model must make their own decisions about data handling, fine-tuning safeguards, and privacy protections. section 3.17
- Privacy oversight still matters. In 2025, European Union regulators launched inquiries into whether DeepSeek, an open-weight language model, had a clear legal basis for processing EU residents’ personal data. section 3.13
Open-weight releases are also a policy choice, not just a technical one. The 2025 US AI Action Plan encouraged them as part of an effort to accelerate innovation. Their practical value depends on access, license terms, and responsible deployment—not on the label “open” alone. section 10.29 section 11.4
