Apertus 1.5 extends the 8B and 70B Apertus 1.0 models with native multimodal understanding (text, images, audio) and an experimental thinking mode, a fourfold longer context window, improved instruction-following, and better tool use.

Key enhancements include native multimodal understanding, a thinking mode for in-depth reasoning, a long context window of 262,144 tokens, improved instruction-following, enhanced tool use, and stronger adherence to the Apertus Charter.

Reproducibility and transparency are emphasized: all recipes, ablations, and data mixtures are documented in the Technical Report, with intermediate checkpoints and curated datasets openly accessible; evaluation methods are designed to be reproducible and auditable by regulators and researchers.

The updated model weights were released to Hugging Face on July 24, 2026 under an Apache 2.0 license, with open weights, open data, open values, and full training details publicly available.

Citations provide formal reference to the arXiv preprint and list extensive author contributions and affiliations associated with the Apertus project.

Pre-training proceeds in four stages (Unpaired, Paired, Cooldown, Long-context) and continues from Apertus 1.0 with refreshed text data and multimodal data meeting open-data and permissive licensing standards; all sources are listed in apertus-data-transparency.

Model availability includes deployments in applications like Proton Lumo and Infomaniak Euria, announcements of API access by multiple providers, and reference to community showcases; comprehensive instructions for running and integrating Apertus 1.5 are provided in model cards and Quickstart guides.

Evaluation now includes multimodal benchmarks, thinking-mode assessments, and extended benchmarks for translation, multilingual capabilities, safety, and tool use, with results reported via EvalCard and the Every Eval Ever (EEE) schema for transparency and reproducibility.

Apertus 1.5 shows strong performance in coding and instruction-following for 8B, competitive multimodal performance closer to Gemma 4 12B among open-weight models, and notable competence in Swiss Legal tasks and Last Translation benchmarks.

The multimodality is implemented via early fusion, tokenizing images with Emu3.5 and audio with WavTokenizer so training remains at text-only speed, with generation still restricted to text.

Apertus 1.5 involves continued pretraining, adding 4 trillion tokens of text and multimodal data for the 8B model and 2 trillion tokens for the 70B model, bringing totals to 19 trillion and 17 trillion tokens respectively.

Post-training comprises five stages (SFT, Long-context SFT, RLVR, DPO, SDPO) to improve reasoning, tool use, and alignment with the Apertus Charter, with released datasets and evaluation pools.