Switzerland releases Apertus, an open-source AI trained on public data

Switzerland has launched an open-source artificial intelligence model called Apertus as an alternative to proprietary systems such as OpenAI’s ChatGPT and Anthropic’s Claude, according to SWI and Engadget reports. The full project is available on Hugging Face, including the source code, training data, model weights, and detailed notes on how the model was built.  The developers say Apertus, whose name comes from the Latin for “open,” is meant to set a new baseline for trustworthy, globally relevant open models. It was trained on material from more than 1,800 languages and arrives in two sizes, with either 8 billion or 70 billion parameters.  Coverage from SWI says its capabilities are comparable to Meta’s 2024 Llama 3 family, signaling performance that could be useful for a wide range of tasks while remaining accessible to researchers and developers. A key goal for the team was compliance. The creators say the dataset followed European Union copyright rules and the voluntary AI code of practice.  To respect website preferences, they limited training to public sources and honored AI crawler opt-out requests, avoiding any stealth collection. That approach aims to reduce legal risk and make it easier for organizations to adopt the model in regulated environments. By publishing the weights and the full data recipe, the project invites outside scrutiny and improvement, allowing companies, labs, and independent developers to fine-tune or evaluate the system for local needs. The two-tier release lets users choose between a lighter 8B model for cost-sensitive or on-premises setups and a larger 70B version for higher quality generation. Positioning Apertus as a transparent, responsibly sourced option also speaks to a broader push in Europe to balance innovation with clear rules, amid criticism from some US companies that tougher regulations could slow AI progress.  With its open license, multilingual training, and documented pipeline, Apertus offers a practical starting point for teams that want modern model performance without opaque data origins, while giving the wider community a shared foundation to test, adapt, and build upon.