Buzzy Ai Multiese makes two smallest models

One of the most famous startup Ai in Europe has released a very small model of AI, he has been named after him after the incident and brain flying.

This multicee computational claim is the smallest model in the world that is still conducted and can handle conversations, speech, and reasonably in one case.

This small honey intended to be mounted to the internet from your device, also open locally on your smartphone, tablet, and PC.

“We can reflect the model to be able to match your device,” says Orús to TechCrunch. “You can run on the premise, directly on your iPhone or at Apple’s watch.”

As previously reported, Multiese computation is the office of Buzzy Ai Restup in the Donostia, Spain, with about 100 employees in the road. It is established co-founded European professors that are the highest of quantum and physics computers, ROMÁn Orús; The quantum the Samuel Muhel. And Enrique Lizaso Ormos, former CEO CEO COO Banc.

It just grows € 189 million (about $ 215 million) In June about the power of the model compression of the model is called “comparatively.” (Due to established in 2019, it has said about $ 250 million.)

Compactifai is an authentic compression algorithm that can reduce the size of an existing model of ai ai without sacrificing the performance of the model, says Orús.

“We have a compression technology that is not a typical compression technology that people are carried out by people of a computer science or studying machine, because we come from quantum physics,” he said. “It is a more subtle compression algorithm.”

This company has dropped a long list of a well-compressed model model, especially the popular small model Calling 4 Scouts or Small mistral 3.1See rankings-. And only launches the compress version of the Two new Opening ModelsSee rankings-. It also compressed a lot of models – offers a slim DeepSeek R1, for example.

But since the business creates a smaller model, it has focused on more to make the smallest model but most powerful.

Two new models are very little to be able to bring the Chat AI capabilities just about IOS devices and work without an internet connection, says the company. It’s hum-gay referring to the family of the Zoo model because of the product’s name is based on the size of the animal’s brain.

The most difficult model is compressed version of the open source model of the SMOLLM2 135. The original has 135m and developed for the use of the device. Superfly is 94m parameters, another for the size of the brain flying. “It’s like flying, but it’s slightly smart,” he said.

Superfly designed to be trained with very limited data, like device operations. Envisions multicee is installed on your Multioned home device, which allows users to open up with voice commands like “starting wash wash” for washing machines. Or the user can ask a problem. With a minor processing of processing (like Arduino), the model can solve the sound interface, as the company shows the techcrunch.

Another model was named Chickbrain, and larger at 3.2 billion parameters, but also more likely and had consideration capabilities. This version Meta Llama 3.1 8B Model, says multicee. But it is small to open MacBook, no internet connection required.

The more important, orús say chickbrain are actually in standard bending, including GSM8K, and GSM8K, and benchmark of Diamond GPQA wisdom.

Here’s the result of the internal test Marta in Chickbrain in the benchmark. This company does not provide a benchmark result for superfly but multiese also does not target the target by using the use of the use.

It is important to note that multiese does not claim that the Zoo model will beat the largest state of art model in the benchmark. Zoo’s performances may not even be in place of leaders. The point is that the technology can shrink the size of the model without hit performance, says the company.

Orús say that the company has been chatting with all the main devices and devices. “We talk to Apple. We talk to Samsung, as well as with Sony and Sony and HP, clearly. HP arrived as an investor in the last round,” he said. The rotation is led by Europe VC Firm Capital Bullhound, with participation of many other people, including HP Tech Ventures and Toshiba.

Startup also offers Tech compressions for other machine shapes, like an image recognition, and in six years earlier clients like Basf, Ally, Moody, Bosch, and Apart.

In addition to selling models directly to the manufacture of major devices, multiese offers A compressed model Through the API held in AWS that can be used by developers, often at the cost of the token below instead of competitors.