One of Europe’s leading AI startups has unveiled two remarkably compact AI models, humorously named after the brains of a chicken and a fly. Multiverse Computing claims these are the tinyest high-performing models available, capable of handling tinquires such as chat, speech, and even reasoning.
The new models are designed for integration into Internet of Things (IoT) devices and can operate locally on smartphones, tablets, and PCs. Orús, a pivotal figure in the company, noted the significance of their compression technology, stating, “We can compress the model so much that they can fit on devices.” This allows for operation directly on platforms like iPhones or Apple Watches, eliminating the necessary for constant internet connectivity.
Headquartered in Donostia, Spain, Multiverse Computing was co-founded by Román Orús, a noted professor of quantum computing and physics, along with experts Samuel Mugel and Enrique Lizaso Olmos, the former deputy CEO of Unnim Banc. Since its founding in 2019, the company has raised around €250 million ($215 million in its latest funding round), largely driven by its innovative model compression technology known as “CompactifAI.”
CompactifAI is a unique, quantum-inspired algorithm that enables the reduction of AI model sizes without compromising performance. Orús describes this compression as “more subtle and refined” than typical methodologies common in computer science or machine learning.
The company has consistently produced compressed versions of popular open-source models, recently launching compact iterations of OpenAI’s latest offerings and extfinishing its work to tinyer models like Llama 4 Scout and Mistral Small 3.1. Despite focutilizing on compression, Multiverse is particularly committed to developing the tinyest, most efficient models.
The newly introduced models, part of a family humorously termed the “Model Zoo,” allow for advanced AI functionalities to be deployed across various IoT devices without the necessity for an internet connection. The first of these models, named SuperFly, is a compressed variant of Hugging Face’s SmolLM2 135, with 94 million parameters—a size comparable to a fly’s brain. It is optimized for restricted data environments, enabling voice command functionalities for home appliances, such as inquireing a washing machine to “start quick wash.”
In contrast, the second model, ChickBrain, boasts 3.2 billion parameters and enhanced reasoning capabilities. It is a compressed version of Meta’s Llama 3.1 8B model, and early benchmarks suggest it slightly outperforms its original in several key assessments. Orús revealed that this model could run effectively on common devices like a MacBook without an internet connection.
It is crucial to clarify that Multiverse does not claim that its models will surpass larger state-of-the-art systems on benchmark leaderboards, but rather aims to demonstrate that tinyer models can maintain a high level of performance without significant loss of quality.
Looking forward, Multiverse Computing is in discussions with leading device manufacturers, including Apple, Samsung, and HP—which also participated as an investor in their recent funding round. Their venture capital round was led by European firm Bullhound Capital, with contributions from various other notable investors.
In addition to its advanced model offerings, Multiverse also provides compression technology for other machine learning applications beyond language processing, having established a diverse client base with companies such as BASF, Ally, Moody’s, and Bosch. Furthermore, Multiverse extfinishs access to its models via an API hosted on AWS, typically at lower fees compared to competitors.
Leave a Reply