OpenWeights

OpenWeights refers to AI models where the trained parameters (weights and biases) are publicly released, enabling local deployment and fine-tuning without full access to the original training code or data.

OpenWeights represents a critical middle ground in AI transparency: the final trained parameters are public, allowing developers to download, run, and fine-tune models on their own infrastructure. This is not full Open Source (which includes training code and data), but it provides significant utility for customization and cost reduction. For example, models like Meta's LLaMA and Mistral 7B exemplify this approach, allowing businesses to fine-tune a 7-billion-parameter model on proprietary data for specialized tasks, often achieving top-tier performance at a fraction of the cost and latency of closed-API systems.

https://huggingface.co/