A new Paris-based research lab wants to make openness part of its approach to artificial intelligence. Kyutai plans to develop foundational models and publish not only the models but also the code and data explaining how they were made.
A nonprofit with substantial backing
Kyutai is a privately funded nonprofit working on artificial general intelligence. It intends to work with PhD students, postdocs and researchers on research papers and open source projects.
At the lab’s announcement, Iliad CEO Xavier Niel said financing had grown to close to €300 million. Niel had initially committed €100 million, and French billionaire Rodolphe Saadé, CEO of shipping and logistics company CMA CGM, is contributing €100 million as well. Other contributors include Eric Schmidt’s foundation and unnamed donors.
The funding is open to further contributions. Saadé said the project could attract other investors, framing the initial backing as a beginning rather than a fixed limit.
Compute and a team for research
Foundational model research requires significant computing resources. Scaleway, Iliad’s cloud division, acquired a thousand Nvidia H100 GPUs, which will be available to Kyutai at cost for model training and inference.
The lab has started hiring its core scientific team. The researchers introduced at the conference were Patrick Perez, Edouard Grave, Hervé Jegou, Laurent Mazaré, Neil Zeghidour and Alexandre Defossez. Their previous experience includes Meta’s AI research team FAIR, Google’s DeepMind division and Inria.
Perez, who previously worked for Valeo, will direct the lab. Scientific advisors Yejin Choi, Yann LeCun and Bernhard Schölkopf will review the team’s work once or twice a year and provide feedback.
Open science means publishing the process
Kyutai’s plan extends beyond releasing open source models. The researchers describe their goal as open science: publishing training source code and data as well, so others can examine how the results were produced.
That approach is also meant to make room for research papers. Niel argued that big tech companies are becoming less tolerant of scientific publications, while publication helps researchers contribute to the wider research community.
The lab does not say speed is its main measure of success. Defossez said Kyutai’s ambition is to provide a scientific purpose, an understanding and a code base that explain its results. The team expects to have something to share within a year.
The researchers also pointed to Mistral AI’s first open source model as a success, noting that community members have fine-tuned it and explored uses based on the Mistral 7B model. Kyutai’s work could likewise give others material to build on, while raising a broader question: whether a research lab can release foundational models more efficiently than private companies, and how those companies might use the results commercially.
France’s debate over AI rules
French President Emmanuel Macron backed open source as a potential French asset in a prerecorded message. He also used the conference to defend France’s position on Europe’s AI Act: regulate uses of AI, he argued, rather than model makers.
Macron said regulation should protect citizens, economic players and democracy while preserving innovation. France has been pushing to water down the AI Act in trilogues, negotiations between the European Parliament, Commission and Council.
Niel supported that position, saying Europe is behind in AI innovation and that regulation could slow newcomers and reduce their chances of catching up. The debate puts Kyutai’s research in a wider context: the lab is pursuing open publication while French leaders argue over how rules can address risks without constraining innovation.