Spanish AI startup Multiverse Computing has released two remarkably tiny AI models dubbed "SuperFly" and "ChickBrain"—named after their brain-sized neural capacities—that can run entirely on-device without internet connectivity.

SuperFly, compressed from 135 million to 94 million parameters, targets IoT applications like smart appliances, enabling voice commands such as "start quick wash" on washing machines. Despite its fly-brain size, it handles chat and speech tasks effectively on minimal hardware like Arduino processors.

ChickBrain, at 3.2 billion parameters, represents a compressed version of Meta's Llama 3.1 8B model that Multiverse claims slightly outperforms the original on several benchmarks, including MMLU-Pro and GSM8K. The compressed model is small enough to run offline on some laptops, like a MacBook. The model demonstrates high-level reasoning capabilities despite being less than half the original size.

Multiverse's quantum-inspired CompactifAI compression technology, developed by quantum physics experts, enables this dramatic size reduction without performance degradation. The company, which raised €189 million in June, is already in discussions with Apple, Samsung, Sony, and HP for device integration.

These ultra-efficient models address growing demand for on-device, offline AI capabilities that preserve user privacy while dramatically reducing cloud processing costs for edge computing applications.