120 #ailights #spiu

A summary of an aiLights webinar hosted by Sabine Wildemann on 22.01.2026 (Recording, Slides). This blog post was written from the transcript with help of Apertus - visit ailights.org for more talks like this one.

My slides are based on the work and initial results of the Apertus team: for full technical details, please read the paper and Tech Report. Special thanks to Verena Kontschieder for the last minute help. I would also like to acknowledge contributions from everyone in the Public AI community for thoughtful exchanges in the past months: Anna, Cole, Jojo, Joseph, Joshua, Lukas, Marco, Natalia, Sam, Tom, Vasco, Wiebke & al.

How do we bring AI from hackathons to reliable public services?

The Swiss Public Inference Utility (SPIU) is a proposed cooperative that aims to support and sustain the mission of Apertus, a Swiss-developed, public-funded, open-source, multilingual large language model. The cooperative should soon be open to individuals and businesses: creating new opportunities to accelerate public AI and the adoption of Apertus and related projects in Switzerland. This is aimed at civil society collaboration with the Swiss AI Initiative, as well as universities, governments, and research centers around the world. The mission is to make AI powered by Apertus more accessible, transparent, and inclusive for public benefit.

Apertus, launched on September 2, 2025, is a large language model that is open data, open source, and open weights. It was trained sustainably in Switzerland using 4096 NVIDIA GPUs over 3 months and is multilingual, supporting languages typically underrepresented in AI models. The model is designed to be legally compliant, reflecting Swiss and European laws, with mechanisms to avoid memorization and filter harmful content. Apertus is not a chatbot in itself, rather a flexible foundational technology that can be used in various applications – e.g.:

• ZüriCityGPT, which aggregrates city data with Apertus or proprietary models
• Dribdat, which supports hackathon teams at events like the Swiss {ai} Weeks
• OpenWebUI, providing a general-purpose, pluggable interface for chatbots

Technical Accessibility

While the initial release required technical expertise to use, the ecosystem has evolved with user-friendly platforms like LM Studio, Ollama, and Open Data Editor. Hardware requirements vary: generally requiring at least 16 GB RAM, which may be a high barrier for cost-conscious or less technically affine users. The Open Web UI used in publicai.co provides a basic chat interface with additional features available with a user account: like file uploads, web page integration, and voice dictation, making it broadly accessible. Such public interfaces allow users to interact with Apertus without needing to download and run the model themselves.

AI as Public Infrastructure

Public AI is a concept that sees AI as essential infrastructure, similar to electricity. It aims to ensure fair universal access, reduce reliance on proprietary models, and promote transparency and accountability. The movement is linked to efforts at Wikimedia and other online communities, supported by research work as well as new platforms like publicai.co: which operate Inference on models from a range of countries that have R&D initiatives like Switzerland. The platform offers both 8B and 70B Apertus models that run on donated compute resources, currently from CSCS, providing free access with rate limits to help ensure fair use.

Building a Sustainable Model

The SPIU cooperative is a new idea with long-reaching roots, that seeks to make public AI infrastructure sustainable through community-based governance, partnerships with other open collectives, and public investment. It plans to address legal and ethical compliance, ensuring that public AI services reflect Swiss values and laws. The cooperative aims to support diverse use cases, including education, small businesses, and startups, by providing a reliable, transparent, and neutral AI infrastructure. We would also work on integrating data sources and APIs (via MCP servers), showcasing useful public AI applications that are integrated with real-world data. Some early ideas of this were shown in the live demo during the presentation. Join the new organization to help to launch SPIU (pronounced sh'pee'oo, like the Swiss-Germanic pronunciation of "play").

Addressing the Performance Gap

One of the key challenges is balancing compliance with performance. Apertus, by design, prioritizes legal and ethical compliance, which can lead to performance disadvantages compared to models that do not respect these constraints. A paper with several members of the Apertus team (Can Performant LLMs Be Ethical? Fan et al, 2025) highlights the Data Compliance Gap, showing that there is a small yet not "significant difference in general knowledge performance between compliant and non-compliant pre-trained models". Continued benchmarking is needed to understand the effect of removing toxic data and respecting copyright, however, this trade-off is already seen as essential for trustworthy and reliable AI services.

Future Developments

Apertus version 1.5 is expected to introduce image capabilities and MCP tooling support, enabling more sophisticated applications. People who are interested in running new releases of model on their own hardware can get some assistance from such a community. The cooperative is also organizing events and meetings to build momentum around Public AI, encouraging everyone interested in the LLM Builder's Summit and AMLD Intelligence Summit in a couple of weeks, and to stay tuned for a launch event for the cooperative at the end of February in Liebefeld, Bern. These efforts overall aim to engage developers, researchers, and businesses to contribute to the Apertus project, and deploy the model in meaningful ways.

How to Get Involved

The cooperative welcomes members from all countries who share its vision of accessible, transparent AI. Technical contributions can be made through open-source development, particularly around apps and knowledge maps of open data. Non-technical users can support the movement by using public AI services, starring the Apertus repository on Hugging Face, participating in community discussions, and providing feedback. The cooperative offers resources and communication channels (like Slack, GitHub, Dribdat and social media) to stay connected and get involved.

The Future of Public AI

The Swiss Public Inference Utility and the Apertus project represent a pioneering effort to make AI a public good, aligning with Switzerland's values of neutrality and inclusivity. By prioritizing transparency, sustainability, and community engagement, the cooperative aims to ensure that AI serves the public interest, supports education and innovation, and maintains high ethical and legal standards. The journey involves navigating technical, legal, and social challenges, but the vision is clear: to build an AI infrastructure that benefits all, not just the privileged few.

The cooperative is not just about providing a service but about fostering a movement in a wider sense that shapes the future of Responsible AI globally, emphasizing collaboration, transparency, and the common good. As the project evolves, it invites the global community to join in, contribute, and benefit from a more inclusive and responsible approach.

Stay tuned for updates on spiu.ch and publicai.ch as the cooperative continues to grow and develop, ensuring that AI becomes a tool for the many, not the few.