Apertus: The Fully Open Foundation Model for Sovereign AI

The Problem with 'Open' AI

In 2026, the term 'open source AI' has become increasingly misleading. Many models labeled as 'open' release only their weights—the trained parameters—while keeping training data, code, and methodology hidden behind corporate veils. This creates significant problems for organizations that need true transparency: regulators who must verify compliance, researchers who need reproducibility, and enterprises that require auditable AI systems.

The EU AI Act demands documentation that most 'open' models simply cannot provide. When you don't know what data trained your model, how can you prove it respects copyright opt-outs? When training code is proprietary, how do you verify safety properties? When alignment methods are secret, how do you audit for bias? These aren't theoretical concerns—they're legal requirements that organizations will need to meet, and current 'open' models leave them without answers.

The situation has created a paradoxical landscape where 'open' models are anything but open in the ways that matter most. Weights without training data are like a baked cake without the recipe—you can taste it, but you can't verify the ingredients, reproduce it, or modify it with confidence. For organizations that need to understand what's inside their AI systems, this level of transparency is insufficient.

Enter Apertus: Fully Open by Design

The Swiss AI Initiative—a collaboration between EPFL, ETH Zurich, and CSCS—has released Apertus, a foundation model that takes 'open' seriously. Everything is transparent: training data, code, weights, methods, and alignment principles. No black boxes, no hidden datasets, no proprietary secrets. The name itself signals the philosophy: Apertus comes from Latin meaning 'open' or 'accessible.' This isn't marketing language—it's a commitment to AI as a public good rather than a proprietary asset.

What makes Apertus genuinely different is the completeness of its release. The training data is published with provenance information, allowing anyone to verify what went into the model. The training code is available on GitHub, including all hyperparameter configurations and optimization choices. The alignment methodology is documented in detail, explaining not just what was done but why. And the weights are released under a permissive license that allows commercial use.

This level of openness is unprecedented in the foundation model space. Even Meta's Llama releases, which are often cited as exemplars of open AI, don't include training data or full training code. Apertus sets a new standard for what 'open' should mean in the context of AI—and that standard is significantly higher than what most of the industry currently meets.

Why This Matters for Sovereign AI

Sovereign AI—the ability for nations and organizations to control their own AI infrastructure—has become a pressing concern. When your AI runs on models you can't inspect, trained on data you can't verify, running on infrastructure you don't control, you've surrendered significant sovereignty. You're dependent on external entities for a critical capability, and you have no way to independently verify that the AI systems you depend on are safe, fair, and compliant with your requirements.

Apertus addresses this directly. Organizations can deploy it on their own infrastructure, audit every component, and modify it to meet specific requirements. For governments subject to the EU AI Act, for healthcare systems handling sensitive data, for financial institutions requiring regulatory compliance—Apertus offers something no proprietary model can: complete transparency and control.

The sovereign AI movement has been gaining momentum globally. Countries including France, Germany, Japan, and India have all expressed interest in developing domestic AI capabilities that don't depend on US or Chinese tech giants. Apertus provides a foundation for these efforts that's both technically credible and genuinely transparent. Instead of trying to build national AI champions from scratch—which requires enormous capital and talent investment—countries can adopt and adapt Apertus to build sovereign AI capabilities on a solid, open foundation.

Technical Specifications and Capabilities

Apertus is a serious foundation model, not a research toy. It's a 70B parameter transformer trained on 2 trillion tokens of carefully curated data. The training data includes web text, books, scientific papers, and code, with detailed documentation of data sources, filtering criteria, and deduplication methods. This allows users to understand exactly what the model was trained on and assess potential biases or gaps.

The model achieves competitive performance on standard benchmarks—comparable to Llama 3.1 70B and approaching GPT-4-class capabilities on many tasks. But the real value isn't in benchmark scores; it's in the ability to understand and modify the model. Researchers can study the training process, identify and fix biases, and contribute improvements back to the community. Enterprises can fine-tune for their specific use cases with full knowledge of the base model's properties.

The training infrastructure is also documented, including details about the compute cluster at CSCS (the Swiss National Supercomputing Centre), the training timeline, and the energy consumption. This level of operational transparency is unheard of in the AI industry and sets another precedent that other labs will be measured against.

The Regulatory Landscape

The EU AI Act is the most prominent example, but it's not the only regulatory framework demanding AI transparency. The US executive order on AI safety, Canada's AIDA, Brazil's AI regulation bill, and various other national frameworks all include provisions that require some level of transparency about how AI systems are trained and deployed. Apertus is uniquely positioned to meet these requirements.

For organizations operating under the EU AI Act, Apertus provides the documentation and transparency needed for compliance. The training data documentation satisfies requirements about data provenance and bias assessment. The open training code allows for independent safety audits. The clear licensing enables commercial deployment without legal uncertainty. No proprietary model can offer this combination of capability and compliance readiness.

The regulatory landscape is still evolving, and requirements will likely become more stringent over time. Organizations that adopt Apertus now are positioning themselves for compliance with future regulations, not just current ones. The ability to audit and modify the model means that as new requirements emerge, the model can be adapted to meet them—something that's impossible with closed proprietary models.

Comparison with Other 'Open' Models

How does Apertus compare to other models that are commonly described as open? The differences are significant and instructive. Meta's Llama series releases weights but not training data or full training code. Mistral's models vary in their openness, with some releases being more transparent than others. Qwen and DeepSeek provide weights with some documentation but fall short of full transparency. Even EleutherAI's models, which are more open than most, don't match Apertus's level of documentation.

Apertus distinguishes itself through the completeness of its release. It's not just that individual components are open—training data, code, weights, methods—but that they're released together as a coherent package with documentation that explains how everything fits together. This makes it possible to actually reproduce the training process, which is the gold standard for scientific openness. Reproducibility isn't just about having the same ingredients; it's about having the recipe and the technique.

Community and Ecosystem

The Swiss AI Initiative has built a community around Apertus that includes academic researchers, industry partners, and government agencies. This community is essential for the model's long-term viability—open source projects succeed or fail based on the strength of their communities, and AI models are no exception.

The community contributes in multiple ways: improving the training data through better filtering and curation, developing new alignment techniques, creating fine-tuned variants for specific use cases, identifying and fixing biases, and extending the model to new languages and domains. This collaborative approach means that Apertus will continue to improve over time, driven by the collective effort of its community rather than the resources of a single company.

For organizations considering adoption, the active community is a significant advantage. It means the model will be maintained and improved, security issues will be identified and fixed, and there's a pool of expertise to draw on for implementation support. This addresses one of the common concerns about open source AI—that it lacks the support infrastructure of commercial offerings.

Challenges and Limitations

Apertus isn't without challenges. The 70B parameter size, while competitive, is smaller than the largest proprietary models. Running it requires significant hardware—though less than larger models—and organizations need appropriate infrastructure. The model's performance, while strong, may not match GPT-4 or Claude on all tasks, particularly those requiring extensive fine-tuning by the original developers.

There are also questions about long-term sustainability. Training foundation models is expensive, and the Swiss AI Initiative will need ongoing funding to develop future versions. The current model was trained on CSCS supercomputing infrastructure, and continued access to such resources isn't guaranteed. The community model of development, while philosophically aligned with the project's values, may struggle to match the pace of well-funded commercial labs.

Finally, openness itself can be a double-edged sword. Full transparency about training data and methods means that the same information that enables auditing and improvement also enables adversarial analysis. Bad actors could study the model's training process to identify vulnerabilities or craft inputs that exploit known weaknesses. This is an inherent tension in open AI that the community will need to navigate thoughtfully.

The Path Forward

Apertus represents a proof of concept for truly open foundation models. Its success or failure will influence whether the AI industry moves toward genuine transparency or continues to use 'open' as a marketing label while keeping the substance closed. The early signs are encouraging—adoption is growing, the community is active, and the model is being used in production by organizations that need the transparency it provides.

For organizations evaluating AI systems, Apertus offers a new option: a capable foundation model that doesn't require you to take anyone's word about what's inside. You can inspect it, audit it, modify it, and deploy it with confidence. In a world where AI is becoming infrastructure, that kind of transparency isn't just nice to have—it's essential.

The open AI movement has been waiting for a model like Apertus. Now that it exists, the question is whether the industry will support it. If enough organizations adopt, contribute to, and build on Apertus, it could become the foundation of a genuinely open AI ecosystem—one where transparency is the default, not the exception.