Swiss AI released the multimodal model Apertus v1.5 with 70 billion parameters

Swiss AI company presented Apertus-v1.5-70B — a new multilingual model with 70 billion parameters. The release is focused on working with complex data and supports extended context up to 262 144 tokens.

The model's architecture provides full multimodality, allowing it to process not only text queries, but also input data in image and audio formats. One of the key features is the integration of a special "thinking mode" designed for performing complex logical tasks and in-depth data analysis.

Developers have ensured the model's compatibility with popular deployment tools, including vLLM and the Transformers library. In internal tests, the model demonstrated a score of 90 points, achieving this result in 1 hour of computation.

The model is distributed as a fully open solution, allowing researchers and developers to integrate it into their own systems without the limitations of proprietary APIs. Using Apertus-v1.5-70B expands the capabilities of local deployment of heavy multimodal systems that require high accuracy of reasoning and work with long contextual windows.