Open-weight means you can download the model itself, not just call an API.
DefinitionWhat it means
An open-weight model is one whose trained parameter weights are published for anyone to download, inspect, fine-tune, and self-host, in contrast to a closed, API-only model where only hosted access is available. Examples include Meta's Llama family, Mistral's models, and DeepSeek's releases; note that open-weight is distinct from fully open-source, since training data and code are often still withheld.
Why it mattersWhy you should care
Open-weight models matter for teams with data-residency, latency, customization, or cost requirements that a hosted API cannot satisfy, since self-hosting removes per-token API fees and vendor lock-in at the price of owning the infrastructure and operational burden. Licensing terms vary widely between open-weight releases, so procurement and legal review of the specific license is a standard step before adopting one commercially.
At a glanceSee it
Open-weight sits between closed API models and fully open-source releases — you get the weights to run, but often under a license that still limits use.
Owning the weights is not the same as needing to host them — the real fork is whether your data or your load justifies running your own GPUs.
Where you see itIn the wild
- Model hubs like Hugging Face hosting downloadable weight files
- Procurement review of an open-weight model's license terms
- Self-hosted inference setups for data-residency compliance