One of Europe’s most important newly established AI companies has released two AI models that are so tiny, named after a chicken brain and the fly brain.
MultiVerse computing claims These are the smallest models in the world that are still high performance and can handle conversation, speech and even reasoning in one case.
These new tiny models are intended to integrate on the Internet devices, as well as run locally on smartphones, tablets and computers.
“We can squeeze the model so much that they can fit into devices,” Orús told TechCrunch. “You can run them in facilities, directly on your iPhone or your Apple Watch.”
As mentioned earlier, MultiVerse Computing is a buzzy European AI start -up to Donostia, Spain, with about 100 employees in offices worldwide. Co -founded by a leading European Professor of Quantum and Physics, Román Orús. Expert Quantum Computing Samuel Mugel. and Enrique Lizaso Olmos the former Deputy Managing Director of Unnim Banc.
It simply increased € 189 million (about $ 215 million) in June with the power of a model compression technology called “compacttifai”. (Since it was founded in 2019, it has raised about $ 250 million, Orús said.)
Compactifai is a compression algorithm inspired by quantum that reduces the size of existing AI models without sacrificing the performance of these models, Orús said.
TechCrunch event
Francisco
|
27-29 October 2025
“We have a compression technology that is not the standard compression technology that people will make from computer science or mechanical learning because we come from quantum physics,” he described. “It’s a thinner and more sophisticated compression algorithm.”
The company has already released a long list of compressed open source models, especially popular small models such as the Llama 4 Scout or the Mistral Small 3.1. And it just started compressed versions of Openai’s two new open models. It has also compressed some very big models – it offers a Deepseek R1 Slim, for example.
But since it is in the business to make models smaller, it has focused particularly on the construction of smaller but more powerful models.
Its two new models are so small that they can bring AI Chat Ai capabilities on any IoT device and work without internet connection, the company says. It calls this family to the zoo model because it names products based on animal brain sizes.
A model that calls Superfly is a compressed version of the Hugging Face Smollm2 135 open source model. The superfly is 94 m. “This is like having a fly, but a little smarter,” he said.
Superfly is designed to be trained in very limited data, such as the functions of a device. MultiVerse envisions that it is incorporated into home appliances, allowing users to handle them with voice commands such as “Start Quick Wash” for a washing machine. Or users can ask troubleshooting questions. With a little processing power (such as an arduino), the model can handle a vocal interface, as the company showed in a live demo in TechCrunch.
The other model is called ChickBrain and is larger in 3.2 billion parameters, but is also much more capable and has reasonable potential. It is a compressed version of Meta’s Llama 3.1 8B model, says Mulrtagee. However, it is small enough to run on a macbook, no internet connection is required.
Most importantly, Orús said that ChickBrain actually exceeds the prototype in various patterns, including MMlu-Pro, mathematical right right right right right points 500 and GSM8K and GPQA Diamond.
Here are the results of ChickBrain’s internal tests at the benchmarks. The company did not offer reference results for the Superfly, but the multiiverse also does not aim in cases of use that require reasoning.
It is important to note that MultiVerse does not claim that its model zoo will win the largest models of state -of -the -art in such points of reference. Zoo performances may not even land on leaderboards. The point is that its technology can shrink the size of the model with no performance hit, the company says.
Orús says the company is already in talks with all the top device and device manufacturers. “We’re talking to Apple. We’re talking to Samsung, also with Sony and HP. Obviously, HP came as an investor in the last round,” he said. The round was driven by a well -known European company VC Bullhound Capital, involving many others, including HP Tech Ventures and Toshiba.
Startup also offers compression technology for other forms of mechanical learning, such as image recognition, and in six years has received customers such as BASF, Ally, Moody’s, Bosch, and others.
In addition to selling his models directly to large device manufacturers, Mulrterse offers the compressed models Through an API hosted on AWS that can be used by any programmer, often at lower pay than competitors.
We always try to evolve and by providing some image of your perspective and feedback on TechCrunch and our coverage and events, you can help us! Complement this investigation To let us know how we do and have the opportunity to win a prize in return!
