Apertus AI: The Swiss Open LLM for Sovereign AI

Apertus model explained: a Swiss LLM from EPFL, ETH Zurich and CSCS

EPFL, ETH Zurich and CSCS released Apertus, an open language model, on 2 September 2025 as part of the Swiss AI Initiative. In ETH Zurich's words, it is Switzerland's first large-scale, open, multilingual language model. ETH Zurich pitches it as a model others can build chatbots, translation systems or teaching tools on.

Apertus release timeline from September 2025 to Apertus 1.5

From the 2 September 2025 release to a Proton partnership listed on 22 September 2026, every dated step in the figure comes from the page that announced it. Where a page gives only a day and month, so does the figure.

Apertus timeline from the primary sources

2 Sep 2025 8B and 70B models released (ETH Zurich)

17 Sep 2025 Main paper published on arXiv (Hugging Face)

May 2026 v1.1 distillation paper and the v1.1-4B-Instruct repository (Hugging Face)

24 Jun Apertus Mini collection updated, 17 items (Hugging Face)

24 Jul Project site lists Apertus 1.5, headline only (apertvs.ai)

22 Sep 2026 Partnership with Proton's Lumo AI assistant listed (CSCS)

Items without a year come from pages that print only the day and month.

The three institutions first announced the joint project in July 2025, two months before launch. Since then, smaller v1.1 models, a second paper and an Apertus Mini collection on Hugging Face have followed.

The project site has an Apertus 1.5 entry dated 24 July that reads "Multimodal input, reasoning, longer context, and better instruction following" and nothing more. The CSCS news list has a 22 September 2026 entry saying the Apertus team is working with Proton on its Lumo AI assistant.

Swiss open source LLM and sovereign AI: what Apertus publishes

For Apertus, "fully open" covers the training data, code, weights, methods and alignment principles, all documented and reproducible, not just the weights. A weights-only release would leave out the data and code that produced them. ETH Zurich's list for Apertus goes further: documentation and source code for training and for the datasets, plus model weights with intermediate checkpoints, all under the same permissive license.

The v1.1 model card also includes scripts for rebuilding the training data. The project site's slogan puts it this way: "Apertus is to AI as Open is to Source."

The site's tagline links that openness to sovereign AI, calling Apertus a fully open foundation model for sovereign AI. Martin Jaggi of EPFL says the release is meant as a blueprint for building a trustworthy, sovereign and inclusive AI model. None of the sources defines sovereignty, but they do say that public institutions built the model and documented how.

Apertus 8B, Apertus 70B and the smaller v1.1 models: sizes and specs

Apertus comes in two generations, the 8B and 70B models trained on 15 trillion tokens and the v1.1 models from 0.5B to 4B, which were distilled from the 8B. “Apertus model sizes and training” compares their sizes and training.

Apertus model sizes and training

In distillation, a small student model learns to copy a larger teacher. For v1.1, the card describes a loss that mixes KL-divergence (how far the student's output distribution sits from the teacher's) with label cross-entropy, 90% to 10%. The 1.7T tokens come from Phase 5 of the original data pipeline, the top quality tier of filtered documents, code and instruction samples, with no new data sources or licenses added.

Swiss AI Apertus training: 15T tokens in over 1,000 languages

Of the 15 trillion tokens behind the 8B and 70B models, spread across more than 1,000 languages, 40% are not English. ETH Zurich says the mix includes languages LLMs have mostly left out, such as Swiss German and Romansh. The project site calls the models multilingual from day one, trained on 1000+ languages, and competitive with top open models at the 8B and 70B scale.

ETH Zurich also lists the resources: more than 10 million GPU hours on the "Alps" system at CSCS, plus the ETH Board, with strategic partners such as Swisscom contributing. The language count depends on the document: the press release and the site say more than 1,000 for the original models, while the v1.1 card says 1811 natively supported languages.

Apertus 4B architecture and compute from the model card

The v1.1 4B model is a 24-layer dense transformer decoder with grouped-query attention, trained on 64 GH200 GPUs using 2.0E22 FLOPs in total. “Apertus v1.1 4B architecture and training” repeats the architecture and training rows from the card.

Apertus v1.1 4B architecture and training

With grouped-query attention, several query heads share one set of key and value heads, which explains the 24 and 8 in “Apertus v1.1 4B architecture and training”. The card also says training documents were packed into 4096-token chunks, with attention across documents masked out.

Apertus LLM benchmark: how the 8B and 4B models score

In the model card's multilingual evaluation, the v1.1 4B Instruct averages 0.473 and the 8B Instruct 0.534. The card calls this a post-training multilingual evaluation against models of similar size, with an Average column next to five task columns: MMLU, TruthfulQA, Arc, IF and LogiQA. It does not explain how the Average is calculated or what IF stands for, so the labels below are copied as printed.

Apertus 4B versus Qwen3, Gemma 3 and EuroLLM on the model card

At 0.473, the 4B Instruct trails Qwen3-4B (0.521), gemma-3-4b-it (0.497), EuroLLM-9B-Instruct (0.480) and SmolLM3-3B (0.479) in the card's own table. Below, the card's columns are split across two tables to keep each one readable.

Benchmark averages for Apertus and similar open models

TruthfulQA, Arc and LogiQA scores for the same models

The 4B's average beats Qwen3-1.7B (0.457), gemma-3-1b-it (0.406) and Qwen3-0.6B (0.401) but falls behind the models in its own size class. On TruthfulQA, though, its 0.506 is the best of the 4B-class rows, ahead of gemma-3-4b-it (0.492), SmolLM3-3B (0.500) and Qwen3-4B (0.497).

No model in “Benchmark averages for Apertus and similar open models” averages higher than the 8B Instruct. Only open models are listed, so that comparison tells you nothing about closed models such as GPT, Claude or Gemini.

Apertus LLM paper: the 2025 report and the 2026 distillation paper

The project is described in two papers, the main report from 17 September 2025 and a distillation paper from 27 May 2026. The main report is titled "Apertus: Democratizing Open and Compliant LLMs for Global Language Environments" (arXiv 2509.14233).

The second, "Apertus LLM Family Expansion via Distillation and Quantization" (arXiv 2605.29128), is by Andrei Panferov, Davit Melikidze, Martin Jaggi and Dan Alistarh. A 6 July news item on the project site also lists the Apertus paper at ACL 2026.

Apertus license and EU AI Act papers: sovereign AI in practice

Apart from its scores, Apertus comes with a permissive license, training data taken only from public sources, and two published EU AI Act documents. The license comes first, since it decides whether you can ship Apertus in a product at all.

License and commercial use: Apache 2.0 on the v1.1 4B model card

The 8B and 70B models carry a permissive open-source license that allows commercial use, according to ETH Zurich, and the v1.1-4B-Instruct repository is tagged apache-2.0. Those two statements are on different pages. The press release never names the license and points readers to Hugging Face for the terms and conditions.

Treat the Apache 2.0 label as applying only to the v1.1-4B-Instruct repository unless the model page you use says the same. Check the license on the exact repository you download.

Opt-outs, PII removal and the EU AI Act documents Apertus publishes

The training data comes only from public sources, is filtered to honor website opt-outs even after the fact, and has personal data stripped out before training starts. ETH Zurich adds that the model was built with Swiss data protection law, Swiss copyright law and the EU AI Act's transparency obligations in mind. The v1.1 model card has a section called "EU AI Act Transparency Documentation and Code of Practice" that links two PDFs, Apertus_EU_Public_Summary.pdf and Apertus_EU_Code_of_Practice.pdf.

The card says the original data summary also covers v1.1, because v1.1 reuses the original release's data in full. Requests to remove personal data or copyrighted content go to the dataset owners or the team, and the card gives the addresses.

If you deploy Apertus, one gap matters. On personal data in model output, the card says "Currently no output filter is provided," and advises checking the site for a filter that reflects deletion requests, then applying it every six months once one exists.

Apertus download options: Hugging Face, Ollama and hosted access

Apertus can be downloaded from Hugging Face and run on your own hardware, or used through Swisscom or the Public AI Inference Utility, the two hosted options named at the 2025 release. Running it yourself gives the most control, while the hosted options spare you the setup.

Hugging Face download: transformers, vLLM and quantized builds

Get a model from the swiss-ai organization on Hugging Face, then load it with transformers 4.56.0 or newer or serve it with the latest vLLM. Quantization stores a model's weights at lower numeric precision so it uses less memory.

For the instruction-tuned models, the v1.1 card offers quantization-aware distillation checkpoints: FP8 and NVFP4A16 aimed at vLLM, and INT3 to INT6 for Apple devices. Its checkpoint table lists the 0.5B, 1.5B and 4B models in BF16 base, BF16 instruct, FP8, NVFP4A16, INT3, INT4 and INT6, while the 8B row has no INT3 or INT6.

For local use, the card points to quantizations for llama.cpp, Ollama and LM Studio and gives a Docker Model Runner command, docker model run hf.co/swiss-ai/Apertus-v1.1-4B-Instruct. For sampling, the card suggests temperature 0.8 and top_p 0.9.

Hosted access through Swisscom and the Public AI Inference Utility

When Apertus launched on 2 September 2025, Swisscom business customers could use it on Swisscom's sovereign Swiss AI platform, and people outside Switzerland could use the Public AI Inference Utility. The press release adds that professionals will find Apertus easy to set up, but that real use needs servers, cloud infrastructure or a user interface on top. No source gives a price for any option, and the details date from the release, so confirm terms and current availability with each provider.

Apertus limits to weigh first, including no PII output filter

Before choosing Apertus, consider three limits: the 4B average trails same-size models in the card's table, there is no output filter for personal data, and Apertus 1.5 is known only from a headline. All three come from the model card and project site cited above.

• Performance: the 4B model's average is below Qwen3-4B and gemma-3-4b-it, and the card says the model may lack capabilities such as advanced math and complex instruction following.
• Personal data: the card says no output filter exists yet.
• Newer versions: the Apertus 1.5 headline lists multimodal input, reasoning and longer context, but the project site gives no size, benchmark or release status.

Weigh these three against the scores above before you decide. If you feed web pages or documents to Apertus, converting them to Markdown first keeps headings and tables as plain text.