Descriptive, not decisive
A flag indicates publisher origin. It is not a statement about training-data origin, cloud region or jurisdiction of every supplier.
A practical layer for teams evaluating where a model comes from, where it can run, what data leaves the EEA and which legal questions remain after self-hosting.
Provider origin describes the primary organization publishing the model. Data residency describes where operational data is processed or stored. Deployment location describes where the inference stack runs. Compliance depends on the concrete legal and technical context.
OpenWeightModels records these dimensions separately so a French, US, Chinese or Canadian model can be evaluated without implying where a European deployer stores prompts, RAG documents, logs or backups.
A flag indicates publisher origin. It is not a statement about training-data origin, cloud region or jurisdiction of every supplier.
Downloadable weights can allow inference on EU/EEA hardware selected by the operator, subject to license and runtime feasibility.
Legal basis, purpose limitation, minimisation, access, deletion, security, processors and international transfers still matter.
Provider and deployer obligations are not determined by a model flag. General-purpose AI rules depend on the actor and use context.
| Layer | EU/EEA question | Typical risk |
|---|---|---|
| Inference server | Where do the weights execute? | External API or remote GPU processing |
| RAG / document store | Where are source documents and chunks stored? | Sensitive content copied to another region |
| Embeddings | Is embedding generation local or external? | Text leaves the selected inference environment |
| Logs / traces | Where are prompts, outputs and traces retained? | Operational logs become a second data store |
| Monitoring / telemetry | Which vendors receive runtime metadata? | Hidden third-party transfers |
| Backups | Where are snapshots replicated? | Residency differs from the primary region |
Permissive licenses such as Apache 2.0 or MIT are generally easier to integrate commercially, but the exact license text, notices, patent terms and any separate use policy remain authoritative. Custom model licenses require a model-specific review.
A technically self-hostable model can still carry contractual restrictions. Conversely, a permissively licensed model can still be deployed in a way that creates privacy, security or regulatory obligations.
The European Commission states that general-purpose AI model providers have documentation, copyright-policy and training-content-summary obligations, with additional obligations for models with systemic risk. The Commission's GPAI guidance entered into application from 2 August 2025.
For downstream organizations, the relevant role can differ. A company merely deploying a model internally is not automatically in the same legal position as the organization placing or significantly modifying a GPAI model on the market.
Record exact model, revision, quantization and runtime. Family names are not enough.
Commercial use, redistribution, fine-tuning, hosted service conditions and notices.
Prompts, RAG, embeddings, logs, monitoring, backups and incident data.
Authentication, network boundaries, encryption, patching and model-server exposure.
Controller/processor relationships, AI Act role, responsible teams and change management.
Model card, license, runtime documentation and dated verification records.
No. Provider origin may matter for procurement and sovereignty strategy, but GDPR compliance depends on the concrete processing and data flows.
Technically yes, when its downloadable weights and the complete runtime stack are operated on EU/EEA infrastructure. Connected services must be assessed separately.
Not automatically. External logging, monitoring, embeddings, backups or support services can still transfer personal data outside the EEA.
No. Weight availability is an access property. Open-source status and any AI Act treatment depend on additional conditions and the relevant legal framework.