# Apertus Apertus is the fully open [[Large Language Models (LLMs)|LLM]] from the Swiss AI Initiative — EPFL, ETH Zurich, and CSCS (the Swiss national supercomputing centre), with Swisscom as strategic partner. The name states the thesis: *apertus* is to AI what *open* is to source. The claim that separates it from "open-weight" models: **everything is open**. Training data, training code, weights, methods, and alignment principles — documented and reproducible. Most so-called open models release weights and stop there; Apertus is one of the few where you can audit what went in. ## Facts that matter - **Sizes**: 8B and 70B parameters, positioned as competitive with open models at equivalent scale - **Multilingual from inception**: 1,000+ languages in training, not an English model patched afterwards - **Privacy engineering in the data pipeline**: crawl opt-outs respected, PII removed, memorization actively suppressed - **Version 1.5** added multimodal input, better reasoning, longer context ## Why it matters This is the credible European answer to [[Sovereign AI]]: public institutions building a reproducible model rather than regulating someone else's. For governments, researchers, and enterprises that need to justify *what the model learned from*, reproducibility is the feature — not benchmark position. It also sets the reference bar for what "open" should mean, against which weights-only releases from [[Mistral AI]], Meta, and the Chinese labs can be measured. The trade-off is honest too: respecting opt-outs and scrubbing PII shrinks the training corpus, so Apertus competes at-scale rather than at-frontier. That's the price of provenance. ## References - [Apertus Website](https://www.apertus-ai.org/) ## Related - [[Large Language Models (LLMs)]] - [[AI Open Weight Models]] — the category Apertus goes beyond - [[Mistral AI]] — the other European model bet, commercial flavor - [[Sovereign AI]] — the concept it implements at national scale - [[Data Sovereignty]] — the legal driver