Multiverse Computing has announced Superfly, a compression AI model that includes just 94 million parameters.
This represents a 15,000-fold reduction compared to the space required for chicken brains, while maintaining the flow of conversation comparable to a much larger AI system.
The small footprint of the model opens up new possibilities for edge computing applications, running locally on virtually any device without an internet connection. Superfly can be embedded in smart appliances for natural language control and reliable AI-assisted vehicles, even in areas with poor network coverage.
While traditional optimization uses techniques such as quantization and pruning, multiverse computing has developed new quantum-inspired methods for achieving breakthroughs.
This development challenges the assumption that advances in AI require vaster models and more stringent hardware, and could move the industry to a more efficient and sustainable AI approach.
Currently, SuperFly is available for private requests, with API access planned in the coming months. This is part of a broader model zoo nanomodel family, including Chickenbrain, a compressed version of Meta’s Llama 3.1 that surpasses the original across multiple benchmarks.
