← Back to feed

The Swiss AI Initiative — a consortium of EPFL, ETH Zurich, and the Swiss National Supercomputing Centre (CSCS) — shipped Apertus Mini on June 15: 16 small language models released to demonstrate distillation and quantization techniques on top of the Apertus foundation. The release follows acceptance of the Apertus technical report at ACL 2026.

The Mini release makes Apertus practically accessible for deployment on constrained hardware. The full suite remains at 8B and 70B parameters, both available under Apache 2.0.

What Apertus Actually Is

Apertus is not a performance-maximised model. It is the most complete attempt yet at a fully auditable, compliance-first foundation for AI in regulated environments:

• 15 trillion training tokens across 1,811 languages, with 40% non-English content
• Retroactive robots.txt compliance: opt-outs filed after the data was scraped are applied backwards against prior crawls — no other open model does this at scale
• Goldfish loss: a pretraining objective that suppresses verbatim data memorisation while preserving downstream task performance
• PII filtering throughout the training corpus
• Full artifact release: training code, data preparation scripts, evaluation suites, intermediate checkpoints, and alignment principles — all permissively licensed

The result is a model purpose-built to meet EU AI Act general-purpose model requirements before they become mandatory.

Benchmarks

Apertus-8B-Instruct scored 31.14% on MMLU-Pro in an independent February 2026 evaluation by DS-NLP Lab — strong for a fully open compliance-first 8B, and behind frontier-class commercial models by a significant margin. The 70B version is claimed to rival or surpass open-weight counterparts on multilingual benchmarks; independent head-to-head comparisons against Mistral Large or Llama 4 Maverick have not been published.

The benchmark gap is partly by design. Goldfish loss, retroactive opt-out compliance, and multilingual prioritisation all shift capability allocation away from English-benchmark maximisation. The trade-off is explicit and documented.

Strategic Partner and Context

Swisscom, Switzerland’s largest telecom, is the first strategic commercial partner. The initiative has already deployed a fine-tuned variant for in-house AI translation in the canton of Ticino.

Apertus is technically Swiss and therefore outside the EU — but inside the European regulatory sphere and the EU AI Act’s reach for models serving EU markets. That positioning matters: it gives the initiative direct compliance standing without the institutional overhead of an EU member-state project.

The model family represents what European sovereign AI actually looks like when executed rigorously: not frontier performance, but reproducible, auditable, legally defensible infrastructure. The Mini release extends that to the edge and to downstream fine-tuning workflows.