Project Apertus released the technical report for its Apertus 1.5 models, as documented by Hugging Face on October 15, 2026. The release updates the earlier 8B and 70B models with image and audio inputs, an optional reasoning mode, and a context window expanded fourfold to 262,144 tokens. Weights for both models were released to Hugging Face on July 24, 2026, under an Apache 2.0 license.
Pretraining added 4 trillion tokens of text and multimodal data to the 8B model and 2 trillion tokens to the 70B model, bringing total training volumes to 19 trillion and 17 trillion tokens. More than half of the training tokens were text-only, drawn from sources including FineWeb2-HQ+, FinePDFs, and Nemotron-CC. The project retroactively respected robots.txt opt-outs across its training data.
Architecture and training pipeline
The team integrated multimodality through early fusion by tokenizing images and audio into the same sequence as text, avoiding a separate vision encoder. Images are tokenized using the Emu3.5 tokenizer, which maps a 512×512 image to 1,024 tokens, while audio is tokenized with WavTokenizer at 40 tokens per second. The models process image and audio inputs but generate only text outputs, and spoken language processing remains experimental.
Engineers continued pretraining across four stages: an unpaired phase with standalone media, a paired phase with image captions and audio transcripts, a cooldown phase with math and code data, and a long-context phase extending the window from 8,000 to 262,144 tokens. Post-training followed five stages: supervised fine-tuning across 3.7 million text samples with multimodal data, long-context fine-tuning, reinforcement learning with verifiable rewards across roughly 350,000 prompts, direct preference optimization, and charter alignment through self-distillation on the 8B model.
Evaluation and deployment
Evaluators tested the models across multimodal benchmarks using the MLLM-eval-suite and language benchmarks on a fork of lm-evaluation-harness. Test results showed the 8B model matched GPT-OSS-20B and outperformed Claude Haiku 4.5 and Google Translate on the Last Translation Benchmark. On Swiss legal tasks, the model scored on par with leading open-weight models.
The project published benchmark records using the Every Eval Ever schema developed with the EvalEval Coalition. This format accounts for 56 of the 60 evaluation reporting requirements specified in Article 55 of the EU AI Act and the General Purpose AI Code of Practice. The reporting infrastructure matches the system used by the UK AI Security Institute to publish open evaluations.
External providers including Swisscom, Infomaniak, and Phoeniqs announced developer API access for the models, while Proton Lumo and Infomaniak Euria deployed them in active applications. Following five million downloads on Hugging Face, the developers announced plans at the AI+X Summit in Zurich for Apertus 2.0, a planned mixture-of-experts model targeting approximately 600 billion parameters.
