Apertus 1.5 technical update details multimodal models and a 2.0 roadmap

The Swiss AI Initiative published Apertus 1.5 release notes on October 15, expanding its account of the models' training and evaluation. The update documents models first released on July 24 rather than announcing their initial availability..

The family includes versions with 8 billion and 70 billion parameters, distributed under Apache 2.0. Both accept text, images and audio while producing text responses, although speech processing remains experimental. They support up to 262,144 tokens of context and offer an optional reasoning mode.

The project has published intermediate model checkpoints, training resources and evaluation records that include testing settings. These materials let researchers examine changes across training stages and reproduce comparisons. The team says performance has improved, while acknowledging that the models still trail leading open-weight systems.

The notes also outline plans for Apertus 2.0, targeting a mixture-of-experts design with approximately 600 billion total parameters. The team identified agent capabilities, mathematics and coding as priorities, without specifying a release date..