s the lifecycle of the models. Section 2 presents the multimodal tokenization, Section 3 the pretraining data and the continued-pretraining recipe, and Section 4 the post-training pipeline. Section 5 reports evaluations across modalities and languages, Section 6 describes the training infrastructure and its efficiency, Section 7 discusses the release and its governance, and Section 8 concludes.
1.1 Released Models, Code, and Datasets
Full openness means every stage of this report can be retraced from released artifacts. We group the release into models, code, and datasets; all code and data carry permissive licenses, and the models are distributed under Apache 2.0.
Models.
The two final post-trained models are distributed on Hugging Face behind a gated download that records assent to the acceptable-use policy and reaches deployers with updates (Section 7.1); major intermediate checkpoints live as branches of these two repositories. The post-training stage checkpoints, one repository per stage under the same license and gating, let the lineage of Section 5.3.2 be inspected checkpoint by checkpoint. Thinking is a request-time flag of the chat template rather than a separate model, even where inference providers expose it as a separate endpoint; usage instructions and supported inference stacks are in the model cards.
swiss-ai/Apertus-v1.5-8B-base
swiss-ai/Apertus-v1.5-8B-SFT
swiss-ai/Apertus-v1.5-8B-RL
swiss-ai/Apertus-v1.5-8B-DPO
swiss-ai/Apertus-v1.5-8B
swiss-ai/Apertus-v1.5-70B-base
swiss-ai/Apertus-v1.5-70B-SFT
swiss-ai/Apertus-v1.5-70B-RL
swiss-ai/Apertus-v1.5-70B
Code.
We release the full training and evaluation stack: the Megatron-LM fork with the multimodal pretraining and SFT extensions, the multimodal tokenizer, the data ingestion inventory, the post-training framework, and the text and multimodal evaluation harnesses. Further special-purpose repositories, such as the tokenizer benchmarks and the synthetic long-context task generator, are linked in the sections where they are used.
swiss-ai/Megatron-LM
swiss-ai/pretrain-code
swiss-ai/apertus-omni-tokenizer
swiss-ai/multimodal-data
swiss-ai/posttraining-data
swiss-ai/posttraining
swiss-ai/self-distillation
swiss-ai/lm-evaluation-harness
swiss-ai/MLLM-eval-suite
swiss-ai/HELMET
swiss-ai/evals-post-train
swiss-ai/model-launch
swiss-ai/verl-recipe
swiss-ai/verl
Datasets.
Apertus is open data. All used training datasets are openly available for download and auditable via the list here swiss-ai/apertus-data-transparency, including reproducible dataset versions and corresponding licenses.
We also release the SFT mixture, the annotated completion pool, and the preference pairs built from it (Section 4.6), as well as the datasets we curate ourselves. The respective sections detail the datasets used in each stage, including exact numbers and the applied preprocessing.
swiss-ai/Apertus-v1.5-SFT-mix
swiss-ai/Apertus-v1.5-Preference-Pool
swiss-ai/Apertus-v1.5-Preference-Data
swiss-ai/nasa
swiss-ai/owid
swiss-ai/smithsonian
swiss-ai/swisstopo
2 Multimodality
One core objective of Apertus 1.5 is to extend the models’ capabilities to understanding and working with images and audio. We train both Apertus 1.5 8B and Apertus 1.5 70B to understand images and audio and respond in text. The multimodal effort spans continued pretraining, supervised finetuning, and long-context supervised finetuning (Figure 1).
This chapter describes how the two new modalities enter the model. Section 2.1 sets out the discrete-token representation and the reasons we prefer it to modality-specific encoders, while Section 2.1.1 and Section 2.1.2 specify the two tokenizers with the vocabulary they contribute. Section 2.2 presents the staged multimodal training data used throughout the pipeline, whereas Section 2.3 describes the tokenization pipelines and their integration into the training infrastructure together with the modifications of the pretraining and SFT codebase to enable multimodal training. Finally, S