Apertus โ Open Foundation Model for Sovereign AI
Apertus: Empowering Sovereign AI via Open Foundation Models
APERTVS.ai represents a breakthrough in Sovereign AI, providing a fully transparent foundation model. This project is the result of a strategic partnership under the Swiss AI Initiative, uniting the expertise of EPFL, ETH Zurich, and CSCS.
๐ ๏ธ Technical Transparency & Access
Apertus is built on the principle of total openness. Unlike proprietary "black box" systems, every component is documented and reproducible:
- Training Data: Full disclosure of datasets.
- Code & Weights: Open access to the underlying architecture.
- Methodology: Detailed training protocols.
- Alignment: Transparent principles for model behavior.
# Example: Accessing Apertus weights
from apertus_lib import load_model
model = load_model("apertus-70b", precision="fp16")
print("Sovereign AI Model Loaded Successfully")
โ๏ธ Regulatory Compliance at Scale
The model is engineered to align with the strict mandates of the EU AI Act. We ensure data integrity through a rigorous pipeline:
- Respects user opt-outs.
- Systematic removal of PII (Personally Identifiable Information).
- Implementation of safeguards to prevent memorization.
Privacy Notice: Your email address will be used exclusively for distributing our newsletter and updates regarding our initiatives. You maintain the right to unsubscribe via the link provided in any email.
๐ Performance Benchmarks
Apertus delivers competitive results when compared to the leading open-source models. The architecture is optimized for two primary scales, where the parameter count is defined as:
| Model Scale | Performance Tier | Primary Use Case |
|---|---|---|
8B | High Efficiency | Edge computing & fast inference |
70B | State-of-the-Art | Complex reasoning & enterprise tasks |
๐ Ecosystem Collaboration
The development of Apertus follows a collaborative flow between Switzerland's top technical institutions:
๐ Latest Updates
Stay informed via our newsletter to track new releases and community research.
- June 15: Apertus Mini โ A collection of 16 lightweight models showcasing advanced quantization and distillation techniques.
- April 25: Academic Recognition โ Our comprehensive technical report was accepted for presentation at ACL 2026.
- March 17: Regional Impact โ Deployment of Apertus for Ticino, a specialized fine-tuned model for internal AI-driven translation.