Apertus 1.5
modelYour notes
The multimodal successor to Apertus, released 24 July 2026 in 8B and 70B sizes. Produced by continued pretraining of Apertus 1.0 — adding a multimodal mix of 4T tokens to the 8B and 2T to the 70B on top of the original 15T — so it keeps the same architecture: a decoder-only transformer with the xIELU activation and the AdEMAMix optimizer, trained exclusively on fully open data.
Three firsts for the family: native multimodal input (image and audio understanding, text output), an optional thinking mode for reasoning tasks, and a 262,144-token context (a 4× jump over Apertus 1.0's 65K). The improved post-training recipe also strengthens instruction-following and tool use. Still Apache 2.0, still positioned as Switzerland's fully open, transparent, EU-AI-Act-aligned sovereign model line.
The promised results arrived with the technical report on 2026-10-11, eleven weeks after the weights — Apertus 1.5: From Open Text Models to Open Multimodal Reasoners (arXiv:2610.15542), 169+ authors across ETH Zurich, EPFL, HES-SO Valais-Wallis, FHNW, SDSC, CSCS, Zurich and Bern. It confirms the architecture is unchanged and that multimodality was integrated through the tokenizer rather than a separate encoder stack, with text-only output retained. The 8B gains up to 20 points on mathematical reasoning in thinking mode, and the family improves on nearly every benchmark against Apertus 1.0, leading the fully open field on knowledge, instruction-following and cultural benchmarks. Models, training recipes, data-preparation details and evaluation suites all ship under permissive licences, trained only on openly available data — the claim the project exists to make. Still not scored on the Artificial Analysis Intelligence Index.
Model Details
Variants
| Name | Parameters | Notes |
|---|---|---|
| Apertus-v1.5-70B | 70B | — |
| Apertus-v1.5-8B | 8B | — |