apertus-70b

Overview. Apertus-70B-Instruct-2509 is a 70.6-billion-parameter dense decoder-only transformer built by the Swiss AI Initiative, a consortium of ETH Zurich, EPFL, and the Swiss National Supercomputing Centre (CSCS), released in September 2025. It's an instruction-tuned model shipped fully in the open — open weights, open training data, and a published training recipe — under an Apache-2.0 licence. Developers who need to audit what a model was trained on, or who want a permissively licensed multilingual base to fine-tune or self-host, would reach for Apertus over closed alternatives.

Key strengths. Apertus's clearest advantage is transparency and multilingual breadth rather than raw benchmark performance: its pretraining corpus spans 15 trillion tokens across 1,811 languages, with roughly 40% non-English content, and Artificial Analysis's openness index scores it at 88.9, reflecting fully published weights, data, and recipe. On general capability benchmarks the picture is more modest — Artificial Analysis's intelligence index sits at 5.1, with GPQA Diamond at 0.27 and IFBench at 0.26, scores well below what's typical for a 70B-class instruction model.

When to use it. Use Apertus-70B-Instruct-2509 when openness and auditability matter more than peak reasoning performance — for instance, when you need to inspect or reproduce the training data and recipe for compliance purposes, or when you want a permissively licensed multilingual base to adapt for a specific language or domain. Its complexity fit is medium, so it suits direct instruction-following tasks better than multi-step reasoning or agentic tool-use workloads, and it has not been verified as safe for sensitive domains such as medical, legal, or financial use.

Limitations. The model's 65,536-token context window is moderate rather than long for current standards, its training data cutoff is March 2024, and it has no extended-thinking or reasoning mode — Artificial Analysis's hard-reasoning benchmarks (GPQA Diamond, HLE) are correspondingly low. It has not been evaluated for safety in sensitive domains such as medical, legal, or financial use.

Sovereignty note. Apertus is built by the Swiss AI Initiative, based in Switzerland rather than the United States, so it does not carry direct CLOUD Act exposure. Because it ships with fully open weights under Apache-2.0, it can also be self-hosted under your own data-residency controls rather than relying solely on a third-party endpoint.