Open Weights
오픈웨이트
The practice of publicly releasing the learned parameters (weights and biases) of a trained AI model, and the parameters themselves. Open weights allow anyone to download, run, fine-tune, and deploy a model on their own infrastructure, but they are distinct from fully open-source AI in that training code, datasets, and intermediate checkpoints are typically not included. The term gained prominence around 2023 with releases such as Meta's Llama series and Mistral, and has since become a standard release mode for major AI labs, valued as a means of securing technological sovereignty against closed API-gated models.
Sources
- Wikipedia (EN) Wikipedia article defining open weights as publicly released learned parameters of trained AI models, distinct from open-source AI.
- opensource.org Open Source Initiative page explaining open weights as final weights and biases of trained neural networks, clarifying their distinction from Open Source AI (no training code, dataset, or full reproducibility).
- hai.stanford.edu Stanford HAI definition: 'An Open-Weight Model is an AI model whose core components are publicly released, allowing anyone to download it.'
- diplomacy.edu DiploFoundation analysis (March 2026) on what open-weight AI means for smaller countries: the ability to run, fine-tune, and deploy models locally without API dependency on foreign companies.
- microsoft.com Microsoft policy paper (July 2026) framing open weights as a foundation for AI accessibility, competition, and institutional sovereignty, modeled on the earlier open-source software movement.