Spanish AI startup Multiverse Computing has released two ultra-compact AI models, named “ChickBrain” and “SuperFly.” The company claims these are the world’s smallest high-performance models, capable of chat, voice, and reasoning. These models are designed to run locally on IoT devices, smartphones, and tablets without the need for an internet connection. The company employs quantum-inspired CompactifAI compression technology to significantly reduce model size without sacrificing performance.
One of Europe’s most renowned AI startups has launched two extremely small AI models, so small they are named after a chicken’s brain and a fly’s brain.
Multiverse Computing claims these are the smallest models in the world that still maintain high performance, capable of handling chat and voice, with one even possessing reasoning capabilities.
These new micro-models are intended to be embedded in IoT devices and to run locally on smartphones, tablets, and personal computers.
“We can compress the models small enough to fit into devices,” founder Román Orús told TechCrunch. “You can run them locally, directly on an iPhone, or even on an Apple Watch.”
As previously reported, Multiverse Computing is a prominent European AI startup based in Donostia, Spain, with offices worldwide and around 100 employees. The company was co-founded by top European quantum computing and physics professors Román Orús, quantum computing expert Samuel Mugel, and former Unnim Bank Vice President Enrique Lizaso Olmos.
The company recently raised €189 million (approximately $215 million) in June, securing investment for its model compression technology called “CompactifAI.” Orús stated that since its founding in 2019, the company has raised approximately $250 million in total.
CompactifAI is a quantum-inspired compression algorithm that can reduce the size of existing AI models without sacrificing performance.
“The compression technology we have is not typical of computer science or machine learning compression techniques because we come from a quantum physics background,” he explained. “It is a more sophisticated and refined compression algorithm.”
The company has already released a long list of compressed versions of open-source models, particularly popular small models like Llama 4 Scout or Mistral Small 3.1. It has also just launched compressed versions of two new open-source models from OpenAI. The company has also compressed some very large models, such as providing a Slim version of DeepSeek R1.
By focusing on making models smaller, the company is particularly dedicated to creating models that are as small yet powerful as possible.
Its two new models are so small that they can bring chat AI capabilities to virtually any IoT device and work without an internet connection. The company humorously refers to this series as a model zoo, as it names products based on the size of animal brains.
One model named SuperFly is a compressed version of the Hugging Face open-source model SmolLM2-135. The original model has 135 million parameters and was developed for edge device use. SuperFly has 94 million parameters, and Orús likens its size to that of a fly’s brain. “It’s like having a fly, but a little smarter,” he said.
SuperFly is designed to be trained on very limited data, such as device operations. Multiverse envisions embedding it in household appliances, allowing users to operate them via voice commands, such as telling a washing machine to “start quick wash” or asking troubleshooting questions. With just a bit of processing power (like an Arduino), the model can handle voice interfaces, and the company demonstrated a live demo to TechCrunch.
The other model, ChickBrain, has a larger parameter count of 3.2 billion but is also more powerful, possessing reasoning capabilities. Multiverse states that this is a compressed version of Meta’s Llama 3.1 8B model. Yet, it is small enough to run on a MacBook without an internet connection.
Importantly, Orús stated that ChickBrain actually performs slightly better than the original model on several standard benchmark tests, including the language skills benchmark MMLU-Pro, the math skills benchmark Math 500 and GSM8K, as well as the general knowledge benchmark GPQA Diamond.
It is worth noting that Multiverse does not claim that its model zoo will outperform the largest state-of-the-art models on these benchmarks. The zoo’s performance may not even make it onto the leaderboard. The focus is on the technology that can reduce model size without compromising performance.
Orús mentioned that the company is already in talks with all leading device and appliance manufacturers. “We are in discussions with Apple. We are talking to Samsung, Sony, and HP, with HP clearly participating as an investor in the last funding round,” he said. This funding round was led by the well-known European venture capital firm Bullhound Capital, with several institutions including HP Tech Ventures and Toshiba participating.
This startup has also provided compression technology for other forms of machine learning, such as image recognition, and has gained clients like BASF, Ally, Moody’s, and Bosch over six years.
In addition to selling models directly to major device manufacturers, Multiverse also offers its compressed models through an API hosted on AWS, which any developer can use, typically at a lower token cost than competitors.
Q&A
Q1: What is special about the CompactifAI compression technology?
A: CompactifAI is a quantum-inspired compression algorithm that differs from traditional computer science or machine learning compression techniques. Since the Multiverse team comes from a quantum physics background, it is a more sophisticated and refined compression algorithm that can significantly reduce AI model sizes without sacrificing performance.
Q2: In what scenarios can the SuperFly and ChickBrain models be used?
A: SuperFly has 94 million parameters and can be embedded in household appliances, allowing users to operate them via voice commands, such as telling a washing machine to “start quick wash” or asking troubleshooting questions. ChickBrain has 3.2 billion parameters and possesses reasoning capabilities, allowing it to run locally on devices like a MacBook without an internet connection.
Q3: Which companies has Multiverse Computing partnered with?
A: The company has established partnerships with several well-known enterprises, including BASF, Ally, Moody’s, and Bosch. It is currently in talks with device manufacturers like Apple, Samsung, Sony, and HP, with HP also participating as an investor in the latest funding round.