Open-Weight Model

A model whose trained weights are publicly available for download, inspection, and modification.

Also known as: open weights, open-source model, open model

In plain English

An open-weight model is like a recipe where the full instructions are shared, not hidden. That means anyone can run the AI on their own computer, check how it was made, or tweak it to better suit their needs.

In practice

Open-weight models let teams host models internally, fine-tune them on private data, and avoid vendor lock-in, which helps with privacy and compliance. They also bring operational costs and engineering work, you need compute resources, monitoring, and careful licence and security review before deployment.

Under the hood

An open-weight model provides downloadable parameter tensors (model weights) and checkpoint files in standard formats (for example, PyTorch state_dict or ONNX). Engineers can perform fine-tuning, quantization, distillation or run on-premise inference, subject to memory, compute and licence constraints; weight format, parameter layout and precision determine compatibility and optimisation options.

Example

Can we use the open-weight model so we can fine-tune it on our customer data and host it internally?

•