EU AI Act
Introduction
Section titled “Introduction”The EU AI Act (Regulation (EU) 2024/1689) governs the provision and use of AI systems in the European Union. This page describes noris’s role within this regulatory framework, as well as the obligations that apply to our customers as operators of their own applications.
noris’s Role: Distributor
Section titled “noris’s Role: Distributor”noris operates the AI infrastructure and provides access to models via standardized API endpoints. Within the EU AI Act, noris takes on the role of distributor. noris is explicitly not the provider (model developer) and not the entity placing the models on the market.
This clear separation is essential for assigning obligations: noris is responsible for infrastructure, API access, and transparency of model selection, while model development remains with the respective provider.
GPAI Classification
Section titled “GPAI Classification”The language models offered on the noris platform are General Purpose AI Models (GPAI). Specialised embedding and reranking models are single-purpose models that do not fall under the GPAI definition (Article 3(63)) and are classified accordingly on their model cards. Under the EU AI Act, the models are classified as follows:
| Attribute | Value |
|---|---|
| Classification | GPAI (General-Purpose AI Model) |
| Risk level | Limited Risk (Article 50) |
| Transparency obligation | Yes |
None of the offered models are classified by noris as high-risk systems. However, classifying a specific use case always remains your responsibility as deployer.
Application and Timetable
Section titled “Application and Timetable”The EU AI Act entered into force on 1 August 2024 and applies in stages:
| Date | Milestone |
|---|---|
| 2 February 2025 | Prohibited AI practices (Article 5) and AI literacy duty (Article 4) |
| 2 August 2025 | Obligations for GPAI models (Article 53), governance, and penalties |
| 2 August 2026 | Full applicability, notably the high-risk regime |
Status of this page: October 2026. All obligations mentioned are therefore in force.
noris’s Transparency Obligations as Distributor
Section titled “noris’s Transparency Obligations as Distributor”As distributor, noris fulfills the following transparency obligations:
- Model cards for all offered models, uniformly structured, with technical specifications, use cases, and limitations
- Links to original sources: HuggingFace pages, license texts, and technical reports from the respective providers
- Traceability of model selection: every model decision is documented and justified
- Infrastructure documentation: locations, security certifications, and data processing are transparently disclosed
- Pass-through of provider information: the model developers’ disclosures (technical documentation, copyright policy, training-content summary) are made available to customers pursuant to Article 53(2) and (3)
Roles and Responsibilities
Section titled “Roles and Responsibilities”| Role | Who | Obligations |
|---|---|---|
| Distributor | noris | Infrastructure, API access, model cards, transparency |
| Provider | Model developer (OpenAI, Google, etc.) | Model documentation, training-data disclosure |
| Deployer | Your company | Use-case risk assessment, AI literacy of staff (Article 4), transparency toward end users, logging |
Upstream Compliance of Model Developers
Section titled “Upstream Compliance of Model Developers”The obligations under Article 53 (technical documentation, copyright policy, training-content summary) rest with the model developers. When selecting models, noris relies on developers that fulfil these obligations and links the original sources on every model card (HuggingFace, licence texts, technical reports). For models whose weights are released under open licences, the derogations under Article 53(3) apply. noris passes on the developers’ information unchanged and does not independently audit training data.
Penalties and Supervision
Section titled “Penalties and Supervision”Violations are subject to fines: up to EUR 35 million or 7 percent of worldwide annual turnover for prohibited AI practices, up to EUR 15 million or 3 percent of worldwide annual turnover for breaches of other obligations. Enforcement lies with the national market surveillance authorities. This page does not constitute legal advice; for the assessment of your specific use case, consult a qualified professional.
Supporting Customer Compliance
Section titled “Supporting Customer Compliance”Limited-Risk Systems
Section titled “Limited-Risk Systems”For limited-risk systems, the central obligation is transparency toward end users (“you are interacting with AI”). Through the OpenAI-compatible APIs, noris provides the building blocks needed to easily integrate these notices into existing interfaces.
High-Risk Systems
Section titled “High-Risk Systems”Should you pursue a use case classified as high-risk, noris supports you with the technical requirements:
- Data governance: vector databases remain under your full control; noris never persists prompts or results.
- Logging: all API requests can be fully logged by you, since noris itself doesn’t persist any content (zero-data-retention).
- Human oversight: the APIs serve as building blocks for human-in-the-loop processes; final decisions remain with humans.
Details on all models and their individual risk assessments are available under Models.
Reference
Section titled “Reference”- Regulation (EU) 2024/1689
