Buzzy AI startup Multiverse creates two of the smallest high-performing models ever

One of the most prominent emerging companies of artificial intelligence in Europe has released two models of artificial intelligence, and they called them a chicken brain and a fly mind.

Multiple computing claims that these are the smallest models in the world that are still highly performance and can handle chat, speech and even thinking in one case.

These new small models aim to integrate into Internet of Things, as well as operate them locally on smartphones, tablets and computers.

“We can press the model to the point that it can fit with the devices,” Urus told Techcrunch. “You can run it in workplace, directly on your iPhone or Apple Watch.”

As we previously informed, the multiple computing is a European start -up company of artificial intelligence in Dunostia, Spain, with about 100 offices around the world. He participated in its founding by the best European professor of computers and quantum physics, Roman Orts; Quantum computing expert Samuel Mughal; And Enrique Lizaso Olmos, former CEO of UNNIM BANC.

I just raised 189 million euros (about $ 215 million) in June, on the strength of the model compression technology called “Compactifai”. (Since its foundation in 2019, it raised about $ 250 million, Orez said.)

Orús said that compactifai is a pressure algorithm inspired by the quantity that reduces the size of the artificial intelligence models that exist without sacrificing the performance of these models.

“We have a pressure technology that is not the typical pressure technology that people will do from computer science or machine learning, because we come from quantum physics,” he said. “It is a more accurate and more accurate pressure algorithm.”

The company has already released a long list of compressed versions of open source models, especially common small models such as Llama 4 Scout or Mistral Small 3.1. Significant versions of new open models were launched from Openai. He also pressured some very large models – it provides Deepseek R1 Slim, for example.

But since they are working in making models smaller, she focused additional attention to making the smallest stronger models possible.

The company says that its two new models are so small that they can bring the ISIS chat capabilities to any Internet of Things and work without connecting to the Internet. Frankly, this family describes this family as a zoo because it is the designation of products based on animal brain sizes.

The model that Superfly calls is a compact version of the Huging Face model. Superfly is 94 meters, which looks like Orús the size of the fly brain. “This is like a fly, but more intelligent,” he said.

Superfly is designed to train in very restricted data, such as device operations. The congress thinks that it is included in the home appliances, allowing users to play them with sound orders such as “Start Washing” washing machine. Or users can ask and repair errors. With a little processing force (such as Arduino), the model can deal with an audio interface, as the company has shown in a live show on Techcrunch.

Another model is called Chickbrain, which is larger at 3.2 billion teachers, but it is also more capable and has thinking capabilities. It is a compact version of the Meta’s Llama 3.1 8B, says Multiverse. However, it is small enough to run on a MacBook, there is no internet connection required.

More importantly, Orús said that Chickbrain is slightly outperforming the origin of the original in many standard standards, including MMLU-PRO, mathematics skills Math 500 and GSM8K standards, and GPQA Diamond.

Below are the results of the internal tests of the Multiverse for Chickbrain on the standards. The company did not provide standard results for Superfly, but Multiver also does not target a super fly in use cases that require thinking.

It is important to note that the multiple universe does not claim that its typical zoo will overcome the largest new models on these standards. The zoo shows may not land on the top boards. The important point is that its technique can reduce the size of the model without hitting the performance.

Orús says that the company is already holding talks with all leading hardware and devices. He said: “We are talking to Apple. We are talking to Samsung, also with Sony and with HP, it is clear that HP came as an investor in the last round.” The tour led a well -known European Capital Capital, with the participation of many others, including HP Tech Ventures and Toshiba.

The start starting also provides a pressure technology for other forms of machine learning, such as images recognition, and in six years I got customers such as BASF, Ally, Moody’s, BOSCH and others.

In addition to selling its models directly to the manufacturers of the main devices, Multiverse offers compressed models through the AWS applications programming interface that any developer can use, often in symbolic fees less than competitors.

We are always looking forward to evolution, and by providing an insight into your view and comments in Techcrunch, coverage and events, you can help us! Fill This survey To inform us how to do and get a chance to win an award in return!