Open weights
Expanded definition
Open weights are not the same as open source. A checkpoint can be downloadable while training data, code, or commercial use is restricted. Teams use open weights for private inference, quantization, and custom serving. You still own evaluation, safety filters, and license review. Hosted APIs that wrap the same family are a different product with different terms. Always read the weight license, not the marketing phrase 'open'.
Open weights vs open source
| Dimension | Open weights | Open source |
|---|---|---|
| Minimum artifact | Downloadable trained parameter files | Source artifacts under an open-source license |
| Training transparency | May omit data, code, and training recipe | Depends on what source components are published |
| Usage rights | Controlled by the model-specific license | Controlled by an approved or source license |
| Deployment | Can support self-hosting when infrastructure permits | Source availability alone does not guarantee deployable weights |
Use the precise term that matches the release. Verify artifacts and license rights instead of treating open weights and open source as interchangeable.
Related terms
Explore adjacent ideas in the knowledge graph.
Open weights FAQ
What is Open weights?
Open weights are downloadable model parameters you may run, inspect, or fine-tune only to the extent the license allows.
How is Open weights used in AI systems?
Open weights are not the same as open source. A checkpoint can be downloadable while training data, code, or commercial use is restricted. Teams use open weights for private inference, quantization, and custom serving. You still own evaluation, safety filters, and license review. Hosted APIs that wrap the same family are a different product with different terms.
Related
Comparisons, tools, and models that connect to this idea.