Open-weight model

With a closed model, you send requests to the provider's API and never see the model itself. With an open-weight model, you can download the weights and run the model locally, on your own servers, or through any hosting provider that supports it. Well-known families include Meta's Llama, Mistral's models, Qwen, DeepSeek, Google's Gemma, and OpenAI's gpt-oss.

Open-weight is not the same as fully open source. The weights may come with a license that restricts some uses, and the training data and code are often not released. Read the license before building a product on one.

The main reasons to choose one are data control, offline or on-premises use, customization through fine-tuning, and cost at high volume. The tradeoffs are running the infrastructure yourself and, for the hardest agent tasks, often lower capability than leading closed models.