Multiverse Computing, based in Donostia, Spain, is turning large language models into featherweights. Using its tensor-network-based compression tool, CompactifAI, the company has created models small enough to run directly on devices such as smartphones, appliances, and cars; no cloud needed. These scaled-down AI models are likened to the brains of chickens or fruit flies, yet retain, or even improve, their reasoning skills, says IEEE Spectrum.
CompactifAI works by scanning a model to identify layers best suited for compression, restructuring them as tensor networks to preserve meaningful patterns, discarding redundant data, then “healing” or fine-tuning the model with light retraining. The results are models that not only shrink but also keep or enhance capabilities.
Two flagship creations illustrate the power of this approach. SuperFly, a 94-million-parameter model (about 30% smaller than its source), is compact enough to run on an iPhone 14 Pro, occupying just 191 MB and handling about 115 tokens per second. ChickenBrain, derived from Llama 3.1 8B, drops to 3.2 billion parameters, i.e., 60% smaller, while outperforming the original on benchmarks such as MMLU-Pro and GSM8K.
Running AI locally brings real advantages: lower latency, fewer privacy risks, and more reliable performance in areas with patchy connectivity; think in-car use through tunnels. Experts note that tensor networks outperform conventional compression methods in many use cases, though more complex reasoning may still pose a challenge.
Multiverse calls these “nano models” and sees fast improvability ahead. Their approach not only makes AI more efficient but it also opens up new possibilities for intelligent edge devices.