Multiverse Computing and Zylon partner to run compressed AI models without internet access

The story so far, in brief; the legwork, and the full article, stay with the original outlet.

Multiverse Computing, based in San Sebastián, and Madrid's Zylon have formed an alliance. Multiverse's compressed AI models will now run natively inside Zylon's platform, which works fully cut off from outside networks. Together they offer a stack that is European from the model to the servers. The two companies disclosed no amount, no timeline and no joint clients.

The problem they are solving is a familiar one. Frontier models usually need large clusters of graphics cards, and those normally sit in someone else's cloud. That is exactly where banks, armies and ministries are not allowed to put their data. Multiverse shrinks models using tensor networks, a mathematical tool borrowed from quantum physics. A bigger model can then run on the server a client already owns, or serve more users at once. Multiverse claims up to 84% better energy efficiency, about 50% lower cost and inference four to twelve times faster. A separate Sopra Steria study of a compressed Llama 3.1 model put energy and CO2 savings at a more modest 30% to 50%.

Zylon built PrivateGPT, an open-source project with over 57,000 GitHub stars. It counts the Spanish Army and the European Commission among its clients. The partners will target defence and edge deployments, public institutions, and regulated sectors such as finance, energy, telecoms, health and manufacturing. "Sovereign AI is not about closing borders," said Enrique Lizaso, Multiverse's co-founder and CEO. His company has raised a round of up to €500 million this year at a €1.5 billion pre-money valuation.