One year in, Switzerland’s open-source AI model Apertus has found real-world uses like translations, but lags behind commercial rivals on coding, math, and reasoning. You can’t train on the murky data everyone else is quietly using in the open. (Sara Ibrahim, swissinfo.ch)
- Switzerland's open-source AI model Apertus hit four million downloads in its first year, with at least 70 known external deployments including canton Ticino using it for administrative translation.
- The model consistently underperforms on coding, math, reasoning, and agentic tasks compared to similarly sized open-weight rivals, and no client-facing deployments have been confirmed by testing firms.
- The core tension is structural: full openness means avoiding the murky training data that commercial labs quietly use, and that gap may be baked in.
