One of the most vital AI startups in Europe has released two AI models that are so small that they called them after the brain of chicken and the brain of flies.
Multiverse computing He claims that these are the smallest models in the world that are still highly efficient and in one case can support chat, speech and even reasoning.
These latest small models are to be embedded in Internet devices, in addition to operate locally on smartphones, tablets and computers.
“We can compress the model so much that they can fit on devices,” said the founding father of Román Orús Techcrunch. “You can start them on the premises, directly on the iPhone or on the Apple watch.”
As we have previously informed, Multiverse computing is a buzzing European startup AI with headquarters in Donostia, Spain, with about 100 employees in offices around the world. He was co -founded by the best European professor of quantum computers and physics, Román Orús; Samuel Mugel quantum calculation expert; and Enrique Lizaso Olmos, former deputy general director of Unnim Banc.
In June, in June he collected 189 million euros (about 215 million dollars) by force of model compression technology, which he calls “Compactifai”. (Since its foundation in 2019, he collected about $ 250 million, said Orús.)
Compactifai is a quantum compression algorithm that reduces the size of the existing AI models without devoting the performance of those models, said Orús.
TechCrunch event
San Francisco
|.
October 27-29 2025
“We have compression technology, which is not a typical compression technology, which is made by people from computer science or machine learning because we come from quantum physics,” he described. “It’s a more subtle and more sophisticated compression algorithm.”
The company has already issued a long list of compressed versions of Open Source models, especially popular small models, akin to Llama 4 Scout or Mistral Small 3.1. And he has just introduced compressed versions of two latest OpenAI open models. He also closed some very large models – for example, he offers Deepseek R1 Slim.
But because he deals with a reduction in models, he focused on additional attention on making the smallest but strongest models.
The company says that two latest models are so small that they’ll introduce AI cap options to any IoT device and work without web connection. He humorously calls this Zoo Zoo family because he calls products based on the sizes of animal brain.
The model he calls Superfly is the compressed version of the Open Source Hugging Face Smollm2-135 model. The original has 135 million parameters and was developed for use on the device. Superfly is 94 million parameters, which ORús compares to the size of the brain of a fly. “It’s like a fly, but a little smarter,” he said.
Superfly is designed for training in the field of very limited data, akin to device operations. Multiverse predicts that he has settled in home appliances, enabling users to handle them using voice commands, akin to “Fast washing” for the washer. Or users may ask questions about problem solving. With low processing power (like Arduino), the model can support the voice interface, as the company showed a live demo for TechCrunch.
The second model is called Chickbrain and is higher at the level of three.2 billion parameters, but it is also much more talented and has the possibility of reasoning. Multiverse says it is a compressed version of the meta lama 3.1 8b model. However, it is sufficiently small to run a MacBook, it does not require an web connection.
More importantly, ORús said that Chickbrain actually barely exceeds the original in several standard comparative tests, including in the scope of the MMLU-PRO, Math 500 and GSM8K reference test and the general knowledge of Diamond GPQA.
Here are the results of internal Multiverse Chickbrain tests on comparative tests. The company didn’t offer comparative results for Superfly, but Multiverse is also not focused on superfly in cases of use that require reasoning.
It must be noted that Multiverse does not claim that his ZOO model overcomes the largest latest models on such comparative tests. Performances in the zoo may not even land on the boards of the leaders. The point is that its technology can reduce the size of the model without hitting performance, says the company.
Orús claims that the company is already in talks with all leading manufacturers of devices and devices. “We talk to Apple. We talk to Samsung, also with Sony and HP, of course. HP came as an investor in the last round,” he said. The round was led by a well -known European company VC Bullhound Capital, with the participation of many others, including HP Tech Ventures and Toshiba.
The startup also offers compression technology for other types of machine learning, akin to recognition of images, and in six years he obtained customers akin to BASF, Ally, Moody’s, Bosch and Others
In addition to selling your models directly for the primary device manufacturers, Multiverse offers Compressed models Through the APS host interface, which might be used by any programmer, often with lower fees for token than competitors.