Model components / Base / pretrained models

Apertus

Switzerland's national open LLM and one of the most transparent large models released to date, built by the Swiss AI Initiative (EPFL, ETH Zurich, and the CSCS supercomputing center) and shipped 2 Sep 2025 in 8B and 70B variants under Apache 2.0. 'Apertus' is Latin for 'open': the release covers not just weights but the full training data (via reconstruction scripts), the Megatron-LM training recipe, intermediate checkpoints, a technical report, and EU AI Act / Swiss data-protection compliance documentation. Trained on 15T tokens spanning 1,811 languages (~40% non-English, including Swiss German and Romansh) on 4,096 GH200 GPUs. Capability is mid-tier, roughly Llama 3.1-70B class on pretraining benchmarks and exceptionally strong multilingually, though below the 2026 frontier on MMLU-style reasoning, but it sets the bar for fully reproducible, legally compliant open models.

Apertus family (8B + 70B, base + Instruct), released 2 Sep 2025 by the Swiss AI Initiative (EPFL, ETH Zurich, CSCS). 15T training tokens across 1,811 natively supported languages (~40% non-English, incl. Swiss German and Romansh), 65,536-token context, trained on 4,096 GH200 GPUs (Alps/CSCS) with Megatron-LM. Verified live on Hugging Face June 2026 (swiss-ai/Apertus-70B-2509).

Openness

5 high confidence

5.0

weights: open(Apache-2.0, BF16 safetensors)
data: open(full pretraining-data reconstruction scripts, github.com/swiss-ai/pretrain-data)
code: open(Megatron-LM training recipe + reconstruction scripts)
checkpoints: open(intermediate checkpoints on HF branches)
license: Apache-2.0(OSI)

Among the most transparent large-model releases to date: open weights, open data (reconstruction scripts rather than a raw dump, to respect opt-out/robots.txt and strip personal data), open training code/recipe, intermediate checkpoints, a technical report (arXiv 2509.14233), and EU AI Act + Swiss data-protection/copyright compliance docs. All Apache-2.0. Sits in the OLMo/Pythia full-openness tier.

https://api.github.com/repos/swiss-ai/pretrain-data recorded 2026-07-30
`private: false`, `archived: false`, license `Apache-2.0`, description "Pretraining data reconstruction scripts for Apertus". The reconstruction scripts are published, which is what satisfies data:open by reconstruction rather than by a released corpus.
https://huggingface.co/swiss-ai/Apertus-70B-2509 recorded 2026-07-30
Model card: "Fully open model: open weights + open data + full training details including all data and training recipes"; "Training Framework: Megatron-LM" linking github.com/swiss-ai/Megatron-LM; "The training intermediate checkpoints are available on the different branches of this same repository"; license apache-2.0; BF16 safetensors.
https://ethz.ch/en/news-and-events/eth-news/news/2025/09/press-release-apertus-a-fully-open-transparent-multilingual-language-model.html recorded 2026-07-30
ETH Zurich press release, title "Apertus: a fully open, transparent, multilingual language model", dated 2 September 2025 (EPFL/ETH/CSCS). Corroborates the release framing; establishes no dimension on its own, since a press release describes rather than exhibits the artifacts.
https://github.com/swiss-ai/pretrain-data recorded 2026-07-30
Repository page for the pretraining-data reconstruction scripts; "reconstruction" appears throughout the tree and README.

Adoption

2 medium confidence

2.0

Newer release (Sep 2025) with strong institutional/sovereign-AI interest but modest raw download volume; swiss-ai/Apertus-70B-2509 shows ~1.5K downloads/month and 149 likes on HF (June 2026); cumulative across the 8B/70B base+Instruct SKUs lands in the 10K-100K band. Distributed via Swisscom, Hugging Face, and the Public AI network. Comparable adoption tier to Falcon 3 / Yi-1.5.

https://huggingface.co/swiss-ai/Apertus-70B-2509 recorded 2026-06-08
~1,462 downloads/month, 149 likes (June 2026)

Capability

3 medium confidence

3.0

Mid-tier capability: competitive with the prior open generation (Llama 3.1-70B class) on pretraining benchmarks and unusually strong multilingually, but below the 2026 frontier on MMLU-style reasoning/knowledge. The value proposition is full openness + multilingual breadth + legal compliance, not raw SOTA, same profile as OLMo 2.

https://huggingface.co/swiss-ai/Apertus-70B-2509 recorded 2026-06-08
pretraining benchmark table: Apertus-70B avg 67.5% vs Llama 3.1-70B 67.3%
https://arxiv.org/pdf/2509.14233 recorded 2026-06-08
Apertus technical report, benchmark results and methodology

Verified 2026-07-30