One of Europe’s most prominent AI startups has released two very small AI models.
Multiverse computing claims that they are still high performance and are the smallest models in the world that can handle chat, speech, and even in one case.
These new little models are intended to be embedded in Internet of Things devices as well as running locally on smartphones, tablets and PCs.
“The more you can compress a model, the more you can fit the device,” founder Román Orús told TechCrunch. “You can run them directly on your property, on your iPhone, or on your Apple Watch.”
As previously reported, Multiverse Computing is a European AI startup headquartered in Donostia, Spain, with around 100 employees in offices around the world. It was co-founded by Roman Ors, the top European professor of quantum computers and physics. Quantum Computing Expert Samuel Mugel; Enrique Lizaso Olmos, former assistant CEO of Unnim Banc.
In June, it raised 189 million euros (approximately $215 million) from the strength of a model compression technology known as “Compactifai.” (Since it was founded in 2019, Orús said it has raised about $250 million.)
Compactifai is a quantum-inspired compression algorithm that reduces the size of existing AI models without sacrificing the performance of these models, Orús said.
TechCrunch Events
San Francisco
|
October 27th-29th, 2025
“We come from quantum physics, so we have compression techniques that are not typical compression techniques that people in computer science and machine learning do,” he explained. “This is a more subtle and sophisticated compression algorithm.”
The company has already released a long list of compressed versions of open source models, particularly the long list of popular smaller models, such as the Llama 4 Scout and Mistral Small 3.1. I also launched two new open models of Openai compressed versions. I also compressed some very large models. For example, it offers DeepSeek R1 slim.
However, as it is a business of minimizing the model, we have focused particular attention on enabling the smallest and most powerful models.
The two new models are so small that they bring Chat AI capabilities to almost any IoT device, allowing them to work without an internet connection. They call this family a model zoo humorously because they name the product based on the size of the animal’s brain.
The model called Superfly is a compressed version of the open source SmollM2-135 model with a hugging face. The original has 135 million parameters and was developed for devices applications. Superfly has a parameter of 94 million, and Orús likens it to the size of the fly’s brain. “It’s like having a fly, but it’s a little more clever,” he said.
Superfly is designed to be trained with extremely limited data, such as device operation. Multiverse is intended to be embedded in home appliances, allowing users to operate with voice commands such as “Start Quick Wash” in the washing machine. Alternatively, users can ask troubleshooting questions. With a little processing power (like the Arduino), the model can handle the Voice interface.
The other model is named Chickbrain, which is large with 3.2 billion parameters, but is much more capable and reasoning. According to Multiverse, this is a compressed version of Meta’s Llama 3.1 8b model. However, it is small enough to run on a MacBook, so no internet connection is required.
More importantly, Orús said that Chickbrain actually outperforms the original on some standard benchmarks. These include the Language Skills Benchmarks MMLU-Pro, the Math Skills Benchmarks Math 500 and GSM8K, and the General Knowledge Benchmark GPQA Diamond.
Below are the results of internal testing of multiverse tickbrain on the benchmark. The company did not provide benchmark results for Superfly, but Multiverse does not target Superfly even in use cases that require inference.
It is important to note that Multiverse does not claim that its model zoo will beat the biggest cutting edge model in such a benchmark. Zoo performances may not even land on the leaderboard. The point is that the technology can reduce model sizes without performance hits, the company says.
Ors says the company is already in talks with all major device and appliance manufacturers. “We’re talking to Apple. We’re talking with Samsung, Sony and HP. The round was led by the well-known European VC company Bullhound Capital, and was attended by many people, including HP Tech Ventures and Toshiba.
The startup also offered compression techniques for other forms of machine learning, such as image recognition, and six years later acquired clients such as BASF, Ally, Moody’s, Bosch and more.
In addition to selling models directly to major device manufacturers, Multiverse offers compression models through AWS hosted APIs that are available to developers, often at a lower token fee than their competitors.
We are constantly aiming to evolve and you can help us by providing insights into TechCrunch and your perspective and feedback on our coverage and events! Fill in this research to let us know how we are doing and get the opportunity to win an award in return!
Source link
