AI Model Provenance & Openness
Why important?
The provenance and openness of models determine the independence and auditability of the AI layer.
How measured?
Scale 0–5 + N/A:
- 0 = Model origin not transparent; only proprietary non-EU models without a switching option.
- 1 = Proprietary models, origin documented, no model switching and no openness.
- 2 = Choice of several models; some open weights; limited portability.
- 3 = Open weights or EU/known model origin; model switching possible without lock-in.
- 4 = Predominantly open or EU-sovereign models; good portability of the AI layer.
- 5 = Full provenance transparency; open-weights or EU-sovereign models; free portability + verifiable evidence.
- N/A = no reliable evidence
Sources / Evidence
Validation questions (RFP)
- Where do the model and its weights come from, in which version, under which licence and with which integrity evidence (hashes, signatures), and is this disclosed per model? Can the model be switched or self-hosted without lock-in? Which AI-specific assessments (e.g. BSI AIC4, ISO/IEC 42001) exist, and with what scope?
Scores comparison
| Provider | Score | |
|---|---|---|
| IONOS Cloud | 4.0 | Open source models (Llama, Mistral) via an OpenAI-compatible API; model choice, no lock-in. (stated) |
| STACKIT | 4.0 | Open-weights models (Llama, Mistral, GPT-OSS) via an OpenAI-compatible API; operated on own K8s platform, open standards, no lock-in. (stated) |
| OVHcloud Public Cloud (inkl. SecNumCloud) | 4.0 | 40+ open-weight models (Llama, Mixtral, Mistral, Qwen, DeepSeek); deployable on-prem/in other clouds; transparent versioning. (stated) |
| Exoscale | 4.0 | Dedicated Inference: any LLM (open models such as Llama, Mistral, GPT-OSS) as a managed, OpenAI-compatible endpoint; no model lock-in. (evidenced) |
| Scaleway | 4.0 | Open-weights models (Llama, Mistral, Qwen3, DeepSeek, gpt-oss) via an OpenAI-compatible API; open ecosystem, no lock-in. (stated) |
| Infomaniak Public Cloud | 4.0 | Open source models, fully hosted in own data centres; model choice per request. (stated) |
| AWS European Sovereign Cloud | 3.0 | Bedrock catalogue mixed: proprietary (Claude, Titan) and open weights (Llama, Mistral); portability workload-dependent. (stated) |
| T Cloud Public | 3.0 | AI Foundation Services: open models (Meta, Mistral) hosted in T Cloud Public, proprietary models (OpenAI, Google, Anthropic) via third-party platforms; mixed provenance. (evidenced) |
| pluscloud open | 3.0 | plusKI: application-specific choice of various LLMs on the open source cloud pluscloud open; early access. (evidenced) |
| noris Sovereign Cloud | 3.0 | AI as a service: shared LLMs (e.g. GPT-OSS-120B) via API plus dedicated managed LLM and customer-owned models; open-weights focus. (evidenced) |
| Cloud Temple Trusted Cloud | 3.0 | LLMaaS: access to current language models (including via IBM watsonx) in a closed SecNumCloud environment; vendor-based model mix. (evidenced) |
| Oracle EU Sovereign Cloud | 2.0 | Model catalogue mixed: Cohere (proprietary) + Llama (open weights); limited choice. (stated) |
| Delos Cloud | 2.0 | Predominantly proprietary OpenAI models (US origin); partly open models via the Azure catalogue. (stated) |
| Microsoft Sovereign Cloud | 2.0 | Flagship proprietary OpenAI models (US origin); catalogue with some open-weights models (Llama/Phi/Mistral). (stated) |
| UpCloud | N/A | No managed model serving; GPU servers for self-hosting, managed Kubernetes announced. Criterion not applicable. (N/A) |
| Hetzner Cloud | N/A | No managed AI or inference offering; GPU servers for self-hosting only. Criterion not applicable to this offering. (N/A) |
| SysEleven OpenStack Cloud | N/A | No managed model serving; AI via self-hosting on MetaKube / OpenStack GPU. Criterion not applicable. (N/A) |