Apertus: a fully open, transparent, multilingual language model (ethz.ch)

EPFL, ETH Zurich and CSCS have released Apertus, a fully open, multilingual large language model whose architecture, weights, training data, recipes and intermediate checkpoints are all published under a permissive open‑source license. Apertus is available in two sizes (8B and 70B parameters), was trained on about 15 trillion tokens spanning over 1,000 languages (≈40% non‑English, including underrepresented tongues like Swiss German and Romansh), and is accessible via Swisscom’s sovereign platform, Hugging Face, and the Public AI Inference Utility. The project includes full documentation, source code and dataset provenance, and enforces data‑use safeguards (publicly available data only, machine‑readable opt‑outs honored, personal data filtered), aligning with Swiss privacy rules and EU AI Act transparency expectations. For the AI/ML community, Apertus is significant because it provides a fully reproducible, large‑scale LLM stack that researchers, engineers and organizations can inspect, adapt and deploy without proprietary black boxes. The two model scales target different use cases (8B for individual/edge usage, 70B for heavier workloads), while the open release and permissive licensing lower barriers for education, research, commercial applications and domain adaptation. Practical deployment still requires compute and tooling, and the team plans iterative updates, efficiency improvements and domain‑specific variants; meanwhile hackathons during Swiss {ai} Weeks will let developers experiment and feed back into future versions.