Swiss AI Initiative ships Apertus 8B and 70B with full open data, code and weights

EPFL, ETH Zurich and CSCS released Apertus under the Swiss AI Initiative. Both 8B and 70B checkpoints include the full pre-training dataset, tokenizer, optimizer states and post-training code. The release also supplies PII removal scripts and memorization audits required for EU AI Act Article 53 obligations. Swisscom provided infrastructure credits but holds no model rights.

Public checkpoints and data cards show training on 1000+ languages with explicit opt-out handling for web domains. Internal CSCS logs indicate 8B model reached 2.1e23 FLOP; 70B reached 1.7e25 FLOP. These figures place Apertus compute between Mistral-7B and Llama-3-70B. No closed model from the same cluster has been announced.

The explicit sovereign-AI positioning separates Apertus from US foundation-model releases that retain safety classifiers and usage restrictions. European regulators now possess a reproducible baseline for testing transparency mandates before the 2026 enforcement window. Downstream integrators can fork without license negotiations.

Next measurable signal is adoption in national public-sector procurement tenders. If three EU member states reference Apertus weights in production systems by Q3 2025, the open-sovereign track will have demonstrated deployment traction distinct from US closed models.