Apertus 8B Instruct is an open-source, instruction-tuned multilingual language model developed by the Swiss AI Initiative, a collaborative effort between ETH Zurich, EPFL, and the Swiss National Supercomputing Centre (CSCS). Released in September 2025, the model is part of a family designed to provide a transparent and sovereign alternative to proprietary AI systems. It is distinguished by its commitment to full transparency, offering open weights, open data pipelines, and documented training recipes.
The model is a decoder-only transformer with 8 billion parameters, pretrained on 15 trillion tokens using the Alps supercomputer. It utilizes several technical innovations, including the xIELU activation function and the AdEMAMix optimizer, which were chosen to stabilize large-scale training. For post-training alignment and instruction following, the developers employed QRPO (Query-Response Preference Optimization) and supervised fine-tuning, focusing on safety and cross-lingual performance.
Apertus 8B Instruct features a context window of 65,536 tokens, supporting long-form document processing and complex agentic workflows. One of its primary strengths is its native support for over 1,000 languages, with roughly 40% of its pretraining data consisting of non-English content. This include underrepresented languages such as Romansh and Swiss German, making it one of the most linguistically diverse models of its size.
Designed with a focus on ethical and legal compliance, Apertus is among the first major models to align with the transparency obligations of the EU AI Act. Its training data respect retroactive machine-readable opt-out requests and prioritize data integrity. The model is released under the Apache 2.0 license, permitting broad commercial, educational, and research use. For inference, the creators recommend using a temperature of 0.8 and a top-p value of 0.9.