Apertus released the 1.5 version of its open-weight AI models on October 15, 2026, introducing native multimodal capabilities and a reasoning-focused "thinking mode" Hugging Face Blog. The update transitions the model family from text-only systems to multimodal reasoners capable of processing images and audio alongside text Hugging Face Blog.
Apertus 1.5 is a continued pretraining of the first-generation models Hugging Face Blog. The 8B model was trained with an additional 4 trillion tokens, bringing its total to 19 trillion, while the 70B model received 2 trillion additional tokens for a total of 17 trillion Hugging Face Blog. The context window was increased to 262,144 tokens, four times the capacity of Apertus 1.0 Hugging Face Blog.
The release includes additions to the 8B and 70B parameter models, which are distributed with open weights, training data, and values Hugging Face Blog. Multimodality is implemented via an early-fusion approach using discrete tokens rather than a separate vision encoder Hugging Face Blog. Images are processed using the Emu3.5 tokenizer, where a 512×512 image equals 1,024 tokens, and audio is handled by WavTokenizer at 40 tokens per second Hugging Face Blog. While the models can understand these modalities, they generate only text Hugging Face Blog.
Apertus claims the 8B model outperforms similarly sized open-weight models on multilingual and instruction-following tasks Hugging Face Blog. The company stated that the 8B model's machine translation capabilities match GPT-OSS-20B and outperform Google Translate and Claude Haiku 4.5 on the Last Translation Benchmark Hugging Face Blog. On Swiss Legal tasks, the models are reported to be on par with leading open-weight models Hugging Face Blog. Multimodal performance is described as competitive in spatial understanding among open-weight models and in document and chart understanding among open-source models, though it remains behind "frontier" open-weight models Hugging Face Blog.
Two applications, Proton Lumo and Infomaniak Euria, have integrated Apertus 1.5 Hugging Face Blog. API access for developers has been announced by providers including Swisscom, Infomaniak, and Phoeniqs Hugging Face Blog. Training recipes, intermediate checkpoints, and data mixtures are available on Hugging Face Hugging Face Blog. The project also published results using the Every Eval Ever (EEE) schema to align with AI governance requirements such as Article 55 of the EU AI Act Hugging Face Blog. Apertus announced a roadmap for version 2.0, which will use a sparse/MoE architecture with approximately 600 billion total parameters Hugging Face Blog.