Definition
An open-weight model, such as Meta's Llama, Mistral's open models, Google's Gemma, DeepSeek or Qwen, can be run without calling the developer's API. That gives control over data, latency and cost, and allows deep customisation through fine-tuning.
Open weights are not always 'open source' in the full sense: training data and code are often not released, and licences can restrict commercial use or large-scale deployment. The US NTIA's 2024 report on open model weights sets out the benefits and risks policymakers weigh.