Swiss AI Initiative Releases Apertus 1.5: Multimodal Open-Weight Models with 262k Context

The Swiss AI Initiative has released Apertus 1.5, an update to its open-weight model family that introduces native multimodal capabilities and extended reasoning features. Built via continued pretraining on the Apertus 1.0 8B and 70B bases, the new models support image and audio inputs, a dedicated thinking mode for improved reasoning, and a context window expanded to 262,144 tokens (4x the previous limit). The models are released under the Apache 2.0 license.

Architecture and Training

Apertus 1.5 uses an early-fusion architecture, processing images and audio as discrete tokens within the same sequence as text, eliminating the need for separate vision encoders. Images are tokenized using Emu3.5 (512x512 images become 1,024 tokens), while audio uses WavTokenizer (40 tokens per second). The training involved adding 4 trillion tokens for the 8B model (totaling 19T) and 2 trillion tokens for the 70B model (totaling 17T), all sourced from publicly available, permissively licensed data.

The post-training pipeline includes five stages: SFT, Long-context SFT, RLVR, DPO, and SDPO. The SDPO stage, which aligns the model with the Apertus Charter, is applied only to the 8B model. The training mix prioritizes text to preserve language capabilities, with over half of the training tokens being text-only.

Performance and Availability

In evaluations, Apertus 1.5 8B outperforms similar-sized open-weight models in instruction-following and multilingual tasks, while thinking mode improves performance on difficult reasoning benchmarks like AIME. Multimodal performance is comparable to Gemma 4 12B, with competitive results in spatial and document understanding. Notably, the 8B model matches GPT-OSS-20B on the Last Translation Benchmark and surpasses Claude Haiku 4.5 and Google Translate in specific translation tasks.

The models are available via Hugging Face and integrated into apps like Proton Lumo and Infomaniak Euria. API access is provided by partners including Swisscom and Infomaniak. The project also announced a roadmap for Apertus 2.0, which will feature a sparse/MoE architecture with approximately 600B total parameters, focusing on agents, mathematics, and coding.