An open-weight model is an AI model whose trained weights are published for download, so anyone can run it on their own hardware under the terms of its licence. Licences range from permissive ones such as Apache 2.0 and MIT to custom community licences with attribution or usage conditions.
Why it matters when licensing company data
Running an open-weight model yourself means the full output, including reasoning, stays on your systems, and no hosted provider's terms of service govern how outputs are used. The model's own licence still applies: some permit commercial use and derivative work explicitly, one model card even names distillation for training other models as permitted, while others add naming, attribution or user-count conditions.
For agent traces this matters twice. Runs on open-weight models can carry the full reasoning, and they avoid the restrictions hosted providers place on using outputs to develop competing models. Check the licence of the exact model version you ran, because terms differ between versions of the same family. Record that version, and where the weights came from, on every run so the check can be repeated later.
Example
A company runs its incident-investigation agent on an open-weight model hosted in its own cloud account, with the weights downloaded under an Apache 2.0 licence. Its traces keep the model's full reasoning between tool calls. When the company reviews whether the traces can be shared outside the business, counsel reads the model licence for that exact version, confirms it places no restriction on using outputs, and turns to the remaining questions: employee notice and customer data in tool results.
Source
DeepSeek-R1 model card (MIT License). Definition reviewed November 9, 2026 by Origin Data Partners.