Apertus 70B Instruct

Apertus 70B Instruct belongs to a fully open language model family developed by the Swiss AI Initiative, positioned around transparency, broad language coverage, and compliance with data-owner opt-outs. The family is released as a 70B and 8B parameter family of decoder-only transformers, with the cataloged variant sized at 70B parameters and trained on roughly 15 trillion tokens across a staged curriculum of web, code, and math data. A defining trait is multilingual depth: the model natively supports over 1,000 languages, making it a fit for cross-language assistants, translation-adjacent workloads, and research scenarios where a single model needs to serve many linguistic communities rather than a handful of high-resource tongues.

The 70B Instruct variant is shaped for general-purpose instruction following on text, reflecting the same architecture and training choices described for the family, including a decoder-only transformer design and post-training via supervised fine-tuning and alignment. Open weights let practitioners self-host and adapt the model for domain-specific assistants, multilingual tools, or integration into pipelines that require local inference. In practical terms, the variant suits teams that prioritize transparency and language coverage over chasing the highest single-benchmark score, and it fits workflows where a fully open data and recipe lineage matters as much as raw capability.