Apertus 1.5 Release Notes

The Swiss AI team behind Apertus has detailed version 1.5 of its fully open language models in a Hugging Face post dated October 15, 2026, with an accompanying technical report. The updated 8B and 70B weights arrived on July 24 under an Apache 2.0 license. The release adds image and audio input, an optional thinking mode, better tool use and instruction-following, and a context window of 262,144 tokens, four times the previous size. Audio understanding is described as experimental, and the models output only text.

Rather than starting over, the team continued pretraining Apertus 1.0, adding 4 trillion tokens to the 8B model and 2 trillion to the 70B, for totals of 19 trillion and 17 trillion. Images and audio are converted into discrete tokens alongside text, avoiding a separate vision encoder and leaving the architecture unchanged. Post-training stages, intermediate checkpoints, data sources and recipes have all been published.

The authors concede the models still trail frontier open-weight systems, but report gains in coding and instruction-following, with the 8B version beating similarly sized open-weight rivals on instruction-following and multilingual tasks. They also cite a translation benchmark where it matched GPT-OSS-20B and outperformed Claude Haiku 4.5 and Google Translate. Results are published in a shared evaluation format developed with the EvalEval Coalition, which the team presents as a way to make scores comparable and useful to regulators.

Proton Lumo and Infomaniak Euria have integrated the models, and Swisscom, among others, has announced API access. The team reports five million downloads and plans an Apertus 2.0 of roughly 600 billion parameters.