Multiverse AI: Tiny, High-Performing Models Revolutionize AI

A Leading European AI Firm Unveils Remarkably Small AI Models

One of Europe's most innovative artificial intelligence companies has recently introduced a pair of AI models distinguished by their exceptionally small size. These models have been playfully named in reference to the brains of a chicken and a fly.

High Performance in a Tiny Package

Multiverse Computing asserts that these represent the world’s smallest AI models capable of maintaining high performance levels. They are designed to handle tasks such as conversational chat, speech processing, and even logical reasoning.

The intended application for these diminutive models is integration into Internet of Things (IoT) devices. They are also engineered to operate directly on smartphones, tablets, and personal computers.

“Our ability to compress these models allows them to function on a wide range of devices,” explained Román Orús, the company’s founder. “They can be run locally, directly on an iPhone, or even an Apple Watch.”

Company Background and Funding

Multiverse Computing, headquartered in Donostia, Spain, is a rapidly growing AI startup with approximately 100 employees globally.

The company was co-founded by Román Orús, a prominent European professor specializing in quantum computing and physics, alongside quantum computing expert Samuel Mugel and Enrique Lizaso Olmos, formerly the deputy CEO of Unnim Banc.

In June, Multiverse secured €189 million (roughly $215 million) in funding, largely due to its proprietary model compression technology, known as “CompactifAI.” Since its inception in 2019, the company has raised around $250 million, according to Orús.

The Power of CompactifAI

CompactifAI is a compression algorithm inspired by quantum principles. It effectively reduces the size of existing AI models without compromising their performance.

“Our compression technology differs from conventional methods used in computer science and machine learning,” Orús stated. “It stems from our expertise in quantum physics, resulting in a more nuanced and refined compression process.”

The company has already released numerous compressed versions of open-source models, including popular smaller models like Llama 4 Scout and Mistral Small 3.1. It has also launched compressed versions of newer OpenAI models and offers a compressed DeepSeek R1 Slim.

Focus on Minimal Size, Maximum Capability

Recognizing the value of smaller models, Multiverse has prioritized developing the most powerful models possible within the smallest footprint.

These new models are small enough to enable chat AI functionality on virtually any IoT device and can operate independently of an internet connection. The company playfully refers to this product line as the “Model Zoo,” naming each model after the brain size of an animal.

Introducing SuperFly and ChickBrain

SuperFly is a compressed version of Hugging Face’s open-source SmolLM2-135 model. The original model contains 135 million parameters and was designed for on-device applications. SuperFly comprises 94 million parameters, a size Orús compares to the brain of a fly. “It’s akin to having a fly, but with enhanced intelligence,” he remarked.

SuperFly is designed for training on limited datasets, such as a device’s operational data. Multiverse envisions its integration into home appliances, allowing voice control for functions like initiating a “quick wash” cycle on a washing machine or providing troubleshooting assistance.

The second model, ChickBrain, is larger with 3.2 billion parameters, but offers significantly greater capabilities, including reasoning abilities. It is a compressed version of Meta’s Llama 3.1 8B model, according to Multiverse. It remains small enough to run on a MacBook without an internet connection.

Notably, Orús highlighted that ChickBrain actually slightly surpasses the original model’s performance on several standard benchmarks, including MMLU-Pro (language skills), Math 500 and GSM8K (mathematical skills), and GPQA Diamond (general knowledge).

Benchmark Results

Here are the results of Multiverse’s internal tests of ChickBrain on the benchmarks.

Multiverse emphasizes that its Model Zoo isn’t intended to outperform the largest, state-of-the-art models on these benchmarks. The primary goal is to demonstrate the ability to reduce model size without sacrificing performance.

Industry Collaboration and Future Outlook

Orús stated that the company is currently in discussions with major device and appliance manufacturers. “We are in talks with Apple, Samsung, Sony, and HP. HP was an investor in our last funding round,” he said. The funding round was led by Bullhound Capital, with participation from HP Tech Ventures and Toshiba.

Multiverse also provides compression technology for other machine learning applications, such as image recognition, and has secured clients including BASF, Ally, Moody’s, and Bosch over the past six years.

In addition to direct sales to manufacturers, Multiverse offers its compressed models through an API hosted on AWS, providing developers with access at competitive token fees.