Spanish AI Startup Multiverse raises $227 million to reduce LLMS and reduce inference costs by 80%

Multiverse Computing, a San Sebastian-based AI startup, has raised 189 million euros (approximately $217 million) to tackle one of the bloated leading language models (LLM), the company announced Thursday

The funding round was led by Bullhound Capital and included participation from HP Inc., Forgepoint Capital and Toshiba. The company says new capital will help expand compression technologies that can reduce LLM.

Multiverse Computing recently introduced a new compression tool called Compactifai, claiming that it can reduce llamas for large language models (LLMS) up to 95% without compromising performance. In reality, that means that businesses can reduce AI-related costs by up to 80%.

A year after development and pilot deployment, the company is ready to scale up with the support of a new round of international and strategic advocates.

Multiverse combines ideas of quantum physics and machine learning to achieve these results, but the technology does not require quantum computers. It is built to mimic how Quantum Systems behaves, but runs on classic hardware.

This latest round has made Multiverse the largest AI startup in Spain, joining the European AI heavyweight ranks such as Mistral, Aleph Alpha, Synthesia, Poolside and Oukin.

The company has already released compressed versions of major open source models such as Llama, Deepseek and Mistral, and plans to add more soon. CEO Enrique Lizaso Olmos says it will focus on optimizing the models companies are already using.

“We’re focusing on compressing the most used open source LLM, which is something companies already use,” Lizasoormos said. “When you go to businesses, most of them use the model Lama family.”

Multiverse tools are already available in the Amazon Web Services AI Marketplace, making it easier for businesses to test and deploy without significantly modifying existing stacks.

How to reduce LLMS bloating to reduce AI costs

The multiverse of core problems deals with: LLM is expensive to run. They usually rely on rugged cloud infrastructures that drive energy bills and limit adoption. Other compression methods, such as quantization and pruning, try to ease the load, but often sacrifice performance in the process.

Compactifai takes a different route. In addition to trimming the model, it also rethinks the structure of neural networks using quantum-inspired techniques known as tensor networks. Results: Smaller, faster, cheaper models that produce the same results. According to Multiverse, its compression model runs 4-12 times faster, reducing inference costs by 50% to 80%.

And it’s not just a cost. These small models are lightweight enough to run not only on cloud or enterprise data centers, but also on local machines such as laptops, smartphones, vehicles, drones, and raspberry PI boards.

“The general wisdom is that we will sacrifice the reduction in LLMS. The multiverse is changing that,” CEO Enrique Lizaso Ormos said. “What began as a model compression breakthrough proved to be rapid and transformative, gaining rapid adoption due to its ability to unlock new efficiencies in AI deployments and radically reduce hardware requirements for running AI models.”

The science behind Compactifai comes from co-founder Román Orús, who helped pioneer the tensor network approach. “For the first time in history, we can profile the inner workings of neural networks, eliminating billions of false correlations and truly optimizing all kinds of AI models,” Ors said.