Right now AI is largely dominated by tech giants, and opaque models. But change is coming. Developed entirely on public infrastructure, Switzerland’s new open-source large language model (LLM) Apertus (Latin for “open”) has now been released. It aims to redefine what open-source AI can be.

The Apertus LLM was trained on the Alps supercomputer by researchers from EPFL, ETH Zurich, and the Swiss National Supercomputing Centre (CSCS). Besides great performance, the Apertus LLM is focused on sovereignty, multilingualism, transparency, and trust.

Looks good on paper, but what’s under the hood?

Apertus: A Sovereign Effort, Public by Design

While Meta’s Llama series, Mistral’s efficiency-focused models, and Falcon’s scale have attracted attention for their technical prowess, all operate within ecosystems formed by commercial pressures or state-affiliated oversight (Deepseek for instance). Apertus, Switzerland’s new open-source LLM stands apart. It is a public initiative from start to finish, rooted in academia, and supported by the Swiss AI Initiative, which is a coalition of over 10 academic institutions backed by the ETH Board and aligned with the European Laboratory for Learning and Intelligent Systems (ELLIS).

Its development shows a deliberate shift toward sovereignty in artificial intelligence. Everything about the model is open: its weights, architecture, and even the training dataset. It’s an approach that starkly contrasts with the walled gardens of GPT-4 or Claude 3.

The idea is simple, fully open models enable high-trust applications and are necessary for advancing research about the risks and opportunities of AI. Transparency in the new LLM is therefor a foundational design choice.

Apertus has Multilingualism at Its Core

Where other models lean heavily toward English or a narrow set of languages, Switzerland’s open-source LLM is trained on data from over 1,500 languages. Apertus includes many languages that have so far been underrepresented in LLMs, such as Swiss German, Romansh, and many others.

Roughly 60% of the corpus is English; the other 40% spans the globe. This approach ensures equitable representation for under-resourced languages and positions the model as the most globally inclusive LLM ever released.

The team was determined to make the models massively multilingual from the start.

The Alps Supercomputer: Powering Public Intelligence

The Apertus model was trained on Alps, one of the world’s most advanced supercomputers. Located in Lugano, the machine is powered by over 10,000 NVIDIA Grace Hopper Superchips and operates entirely on carbon-neutral electricity, cooled by water from Lake Lugano.

Alps delivers 270 petaflops of computing power. That performance – comparable to or exceeding some of the top-tier cloud infrastructures globally – allowed the Swiss team to train two model sizes: an 8-billion-parameter base and a 70-billion-parameter flagship. The larger model puts it in the same league as LLaMA 2 (70B) and Falcon (180B), but with far greater multilingual depth.

How Apertus Stacks Up: Comparisons with Other Open-Source LLMs

Below I will compare the new open-source LLM Apertus with 4 other open-source LLMs.

LLaMA 3 (Meta): Offers various sizes, including a 70B parameter version. Trained on approximately 15 trillion tokens, similar in scale to the Swiss model. However, LLaMA’s license restricts some forms of commercial use and does not disclose training data sources.

Mistral 7B/30B: Built for speed and memory efficiency, Mistral outperforms larger models on many benchmarks. But it’s optimized primarily for English and code tasks. The Swiss model provides broader use cases across multilingual domains.

Falcon 180B: Trained by the UAE’s Technology Innovation Institute, Falcon is one of the largest open-weight models, with 3.5T training tokens. Yet it lacks the same level of transparency and reproducibility in data sourcing as the Swiss model.

BLOOM (BigScience): A milestone in collaborative open-source LLM development. Trained on 46 languages, it champions multilingualism but doesn’t approach the Swiss model’s language diversity. With 176B parameters, BLOOM is much larger, but that comes at the cost of accessibility and inference efficiency.

Swiss LLM Apertus: What sets it apart is the full trifecta: openness, multilingualism, and sovereign infrastructure. No other LLM on the market offers this combination with such transparency and regulatory alignment.

In short, if we compare all 5 LLMs we come to this conclusion:

Feature Apertus Open-Source LLM Other Open LLMs
Openness 100 % transparent (data, code, weights) Varies; Llama open weights but confidential data provenance
Languages Supported 1 500+ languages BLOOM ~46, others mostly multilingual but limited
Parameter Range 8B & 70B Variety (7B–180B)
Compute Infrastructure Alps supercomputer, carbon‑neutral Cloud, commercial HPC; some public
Licensing Apache 2.0 Mixed: Apache, custom, some restrictions
Best Suited For Global research, public sector apps, auditability Dev tools, chatbots, efficiency-focused uses

Regulatory-Ready and Ethically Aligned

Switzerland’s open-source LLM Apertus complies with the EU AI Act and Swiss data protection laws. Researchers demonstrated that opting out of web-crawling during dataset creation results in no measurable performance loss – countering the common argument that consent-based data usage hinders model quality.

This makes the model particularly suited for deployment in government, education, healthcare, and legal domains, where explainability and compliance are non-negotiable.

Because the Swiss open-source LLM was developed using transparent and opt-out-respecting data practices, it may offer a safer and more accountable alternative for academic use than models like ChatGPT, which rely heavily on large-scale web-crawled data of uncertain provenance.

It’s worth noting that academic safety also depends on how language models are used – for example, in avoiding plagiarism, managing bias, and mitigating hallucination risks – not solely on how they were trained. That said, the fact that the Swiss model respects web-crawling opt-outs without sacrificing performance represents a meaningful step forward in responsible data practices.

Criticism of Switzerland’s Open-Source LLM Apertus

Although direct critiques are limited, I did pick up discussions across platforms such as Hacker News and X which reveal a number of recurring themes:

1. Can Apertus Compete with Industry Giants?

A recurring question is whether Switzerland’s open-source LLM can compete with commercial LLMs from OpenAI, Anthropic, Meta, or xAI. While the 70-billion-parameter Swiss model is sizable for an open-source initiative, it lags behind the estimated 300B–1.8T parameters of Claude 4 Opus or GPT-4. Critics doubt whether a publicly funded project – even with access to the Alps supercomputer – can match the performance, scale, and fine-tuning optimization of industry-backed models.

2. Limited Experience in Large-Scale LLM Training

While Swiss academia boasts strong machine learning talent, it lacks deep experience in large-scale LLM training. The argument is that much of the performance advantage in commercial models stems not just from data or architecture, but from mature, infrastructure-driven optimization practices developed over years in U.S. and Chinese tech ecosystems.

3. Infrastructure Dependency and Scalability

The Swiss model’s reliance on the Alps supercomputer – powered by 10,000+ NVIDIA Grace Hopper Superchips – is both a strength and a potential barrier. While Alps enables efficient training at scale, the model’s 70B size may be impractical for smaller institutions or developers lacking comparable infrastructure. This could limit fine-tuning and downstream adoption outside of well-resourced environments.

4. Dependence on Community Engagement

Unlike corporate-backed models with dedicated engineering teams, the Swiss open-source LLM’s future evolution depends heavily on contributions from the open-source community. Critics warn that if this ecosystem fails to gain momentum, the model may stagnate, falling behind faster-moving proprietary alternatives. The permissive Apache 2.0 license supports innovation, but community traction remains uncertain.

5. Concerns About Safety and Open-Source Risks

Despite its emphasis on transparency and compliance with EU and Swiss data laws, some users on X have expressed concern about the safety and trustworthiness of a fully open model. Open-source LLMs can be repurposed or fine-tuned in unintended ways, raising questions about whether sufficient safeguards exist to prevent malicious use, especially given the lack of usage restrictions under permissive licenses.

6. Ethical Trade-Offs and Performance Gaps

The model’s strict adherence to ethical data collection – such as honoring web crawling opt-outs – is praised, but some critics argue this might reduce training data diversity. While project leaders claim this causes “virtually no performance degradation” for general tasks, doubts remain about its performance in niche or high-specialization contexts that depend on broader or domain-specific datasets.

7. Narrow Scope of Application

Unlike general-purpose models like ChatGPT, the Swiss open-source LLM is explicitly designed for socially beneficial applications in science, education, healthcare, robotics, and climate research. While this aligns with Switzerland’s “ProSocial AI” framework, it may reduce the model’s attractiveness for commercial developers or businesses seeking broader AI functionality.

Open Science, Open Source, Global Impact

The Apertus model has been released under an Apache 2.0 license with full documentation, including training methods, architecture details, and usage guidelines. Organizations worldwide will be able to fine-tune, adapt, and integrate it into their workflows – free from vendor lock-in.

The release builds on the momentum of the Swiss AI Initiative, which offers over 20 million GPU hours per year to researchers. It’s the largest coordinated open science effort in the foundation model space globally.

Switzerland’s open-source LLM Apertus is a good counterweight to the deep learning breakthroughs which mostly come from corporate labs guarded by NDAs and restricted APIs .

It may not dominate every benchmark or corner every market, but what it offers is vital: open intelligence, rooted in trust, created by the public, and available to all.

As far as the criticism of Switzerland’s open LLM effort is concerned, it centers on whether it can scale to meet the performance, flexibility, and adoption levels of leading proprietary models. Key concerns include its relative lack of large-scale training experience, reliance on centralized infrastructure, and dependence on community support. While its transparency, multilingualism, and ethical approach mark major advances, critics question if those strengths are enough to offset the power, polish, and pace of commercial alternatives.

A real analysis will follow once the Apertus model has been thoroughly vetted.


I specialize in sustainability education, curriculum co-creation, and early-stage project strategy. At WINSS, I craft articles on sustainability, transformative AI, and related topics. When I'm not writing, you'll find me chasing the perfect sushi roll, exploring cities around the globe, or unwinding with my dog Puffy — the world’s most loyal sidekick.