AI Startup Unveils "Animal Brain" Models for On-Device Intelligence
Multiverse Computing, a rising star in the European AI scene, has unveiled two strikingly compact AI models, playfully named after the respective sizes of a chicken and a fly's brain.
This Spanish startup asserts that these models are among the smallest currently available that still deliver impressive performance. They are designed to handle tasks like text-based conversations, speech recognition, and even, in the case of the "ChickBrain" model, reasoning.
The primary goal is to enable embedding these AI systems directly into everyday objects connected to the Internet of Things (IoT), and for seamless local operation on devices like smartphones, tablets, and personal computers.
"Our compression techniques allow us to shrink these models down to a size that makes them easily deployable on devices," explained Orús to TechCrunch. "This means you can run them offline, right on your iPhone or even your Apple Watch."
Headquartered in Donostia, Spain, Multiverse Computing boasts a team of around 100 employees globally. It was founded by Román Orús, a leading European quantum computing professor, Samuel Mugel, an expert in quantum computing, and Enrique Lizaso Olmos, formerly a deputy CEO at Unnim Banc.
The company recently secured €189 million (approximately $215 million) in funding this past June, driven by the strength of its proprietary model compression technology known as "CompactifAI." To date, it has raised around $250 million since its inception in 2019, according to Orús.
CompactifAI, leverages a quantum-inspired compression approach to reduce the size of existing AI models without significantly impacting their capabilities, states Orús.
"Our compression technology is unlike conventional methods used in computer science or machine learning. Because our origins are in quantum physics, we employ a more nuanced and sophisticated compression algorithm," he said.
Multiverse Computing already offers compressed versions of several popular open-source models, including Llama 4 Scout and Mistral Small 3.1. They have recently added compressed versions of OpenAI's two new open models to their product lineup. The company also provides compressed versions of large language models, such as a DeepSeek R1 Slim.
The company's emphasis is on crafting the smallest, yet most powerful AI models possible, which is key to their business of model size reduction.
The newly introduced models facilitate chat AI on almost any IoT device without the need for an internet connection. Multiverse Computing calls this group of products the "Model Zoo," because the name reflects the size of animal brains.
The SuperFly model, a compressed version of Hugging Face's SmolLM2 135, has 94 million parameters compared to the original's 135 million. Orús humorously described it as "like having a fly, but a little bit more clever."
SuperFly is intended for training on extremely limited data sets, like the operational data of a specific device. Multiverse sees its potential in home appliances, allowing users to control them with simple voice commands. It can also aid in troubleshooting. The company demonstrated this by implementing a voice interface on an Arduino.
ChickBrain, with 3.2 billion parameters, is larger but also demonstrates greater capabilities, including reasoning. Multiverse says it is a compressed version of Meta's Llama 3.1 8B model. Importantly, it is small enough to run on a MacBook without an internet connection.
Even more remarkable, ChickBrain slightly surpasses the performance of the original Meta model on several benchmark tests, including MMLU-Pro, Math 500, GSM8K, and GPQA Diamond.
Multiverse Computing acknowledges that its Model Zoo is not intended to compete with the largest state-of-the-art models in benchmark rankings. The primary focus is the company's ability to reduce model size without sacrificing performance.
Orús revealed that Multiverse Computing is actively engaged in discussions with leading device and appliance manufacturers, "We are talking with Apple, Samsung, Sony, and HP, who are investors in our last funding round." The round was spearheaded by Bullhound Capital, with other investors including HP Tech Ventures and Toshiba.
Beyond specialized AI models, Multiverse Computing also provides compression technology for other types of machine learning, like image recognition. In its six years, it has secured partnerships with companies such as BASF, Ally, Moody's, and Bosch.
In addition to direct sales, Multiverse provides its compressed models through an AWS-hosted API, offering developers potentially lower token fees compared to competitors.
