Apertus: An Open Foundation Model for Sovereign AI

Apertus provides a fully transparent foundation for sovereign AI

Apertus is an open foundation model developed by the Swiss AI Initiative—a collaboration between EPFL, ETH Zurich, and CSCS—designed to enable "Sovereign AI" by making the entire development pipeline public. Unlike many "open-weight" models, Apertus provides open access to training data, code, weights, methods, and alignment principles to ensure reproducibility and transparency.

Technical Specifications and Compliance

Apertus is designed to be competitive with top open models at scales of 8B and 70B parameters. It is natively multilingual, having been trained on over 1,000 languages.

EU AI Act Compliance

A key differentiator for Apertus is its focus on regulatory compliance at scale. The model is built to meet EU AI Act requirements through several specific mechanisms:

  • PII Removal: The model incorporates processes to remove personally identifiable information.
  • Memorization Prevention: Training methods are employed to prevent the model from memorizing specific training data points.
  • Opt-out Respect: The system is designed to respect data opt-outs.
  • Dynamic Filtering: The Swiss AI Initiative (SNAI) provides a downloadable file of hash values that users can apply as an output filter to remove personal data in accordance with data protection deletion requests.

Community Feedback and Performance Analysis

While the vision of Apertus is highly regarded, community feedback from developers and testers highlights a gap between the project's goals and its current performance.

Performance Critiques

Users have reported mixed results regarding the model's actual utility:

  • Multilingual Reliability: Some users report that the model is unreliable for simple linguistic tasks, such as conjugation or translation, noting a tendency to hallucinate non-existent words.
  • Version Maturity: Several contributors noted that Apertus V1 performance was sub-par, and the team is currently working on V2 to address these issues.
  • Instruction Tuning: There are claims that some instruct models may be based on Llama 3.1 fine-tunes rather than entirely original architectures.

Use Case Viability

Despite performance concerns, some users find the model viable for specific implementations:

"It is not ready for agentic use, but as a generic driving model for RAG use cases, it is pretty competent. You can build useful software with it."

The Broader Context of Open Foundation Models

Apertus enters a landscape with other fully open LLMs, such as Allen AI's OLMo 3.1 and MBZUAI's K2 Think V2, which also release full training pipelines and datasets. It stands in contrast to models like Nvidia Nemotron, which provides open training sources but retains some proprietary dataset elements.

The push for "Sovereign AI"—AI developed independently of the dominant US-based commercial labs—is seen by some as a necessary step for data security and technological independence. However, critics argue that the primary hurdle for open AI is not just the availability of models, but the user experience (UX) required to make local LLMs accessible to non-technical users.

Sources