Multiverse Computing Raises $215 Million for AI Model Compression

Multiverse Computing, a Spanish AI startup, has successfully secured $215 million in a Series B funding round. This investment was led by Bullhound Capital, with additional backing from HP Tech Ventures and Toshiba. The company's innovative technology, CompactifAI, is designed to significantly reduce the size of AI models without compromising their performance.
CompactifAI leverages tensor networks from quantum physics to compress AI models, making them small enough to run on smartphones. This technology reportedly reduces the parameter count by 70% and model memory by 93%, while maintaining 97-98% accuracy. The compressed Llama-2 7B model, for instance, runs 25% faster at inference and uses 70% fewer parameters, with only a 2-3% drop in accuracy.
The compression process involves folding weight matrices into smaller, interconnected structures called Matrix Product Operators. This method preserves meaningful correlations while discarding redundant patterns, allowing for dramatic size reductions. After compression, the models undergo a brief "healing" process, which involves retraining that takes less than one epoch. This restoration process is claimed to run 50% faster than training the original models due to decreased GPU-CPU transfer loads.
Multiverse Computing's technology has the potential to address one of AI's most significant challenges: the need for specialized data centers to operate massive models. By making AI models more compact and efficient, the company aims to bring AI benefits such as enhanced performance, personalization, privacy, and cost efficiency to companies of any size.
The company already serves over 100 clients, including Bosch and the Bank of Canada, applying their quantum-inspired algorithms to various fields such as energy optimization and financial modeling. The Spanish government co-invested €67 million in March, pushing total funding above $250 million. Multiverse Computing currently offers compressed versions of open-source models like Llama and Mistral through AWS and plans to expand to other reasoning models such as DeepSeek R1.
HP Tech Ventures' involvement in the funding round signals interest in edge AI deployment, which involves running sophisticated models locally rather than on cloud servers. This technology promises not only cost savings but also enhanced performance, personalization, privacy, and efficiency for companies of all sizes.

Comments
No comments yet