Definition
Open-weight models differ from fully open-source AI because only the final learned parameters are released, not necessarily the infrastructure or data used to create them. This allows users to run inference and fine-tune the model locally without access to the original training pipeline. While it promotes accessibility and reduces barriers to entry, it limits full reproducibility and deep architectural understanding compared to fully open-source initiatives.
Summary
AI models where the trained parameters (weights) are published, but the training code and dataset may remain private.
Key Concepts
- Model weights
- Inference
- Fine-tuning
- Partial transparency
Use Cases
- Deploying LLMs on local hardware
- Customizing pre-trained models
- Reducing API costs