Europe’s most prominent AI startup has released two AI models that are so small, they have named them in the name of Chicken’s brain and bees.
Multiwesi computing Claims are the smallest models in the world that are still high -performing and can handle chat, speech, and even arguments in one case.
These new small models aim to be embedded in the Internet -off -item equipment, as well as running locally on smartphones, tablets and PCs.
“We can compress the model so much that they can fit on the devices,” founder Roman and Roman told Tech Crunch. “You can run them directly on your iPhone, or on your Apple Watch, in the premises.”
As we have already reported, multi -virus computing is an elderly European AI startup located in Donostia, Spain, with close to 100 employees at offices around the world. It was founded jointly by a top European professor of quantum computers and physics. Samuel Magel, a quantum computing specialist; And Enrik Lizaso Olymos, a former deputy CEO of Anim Bank.
It collected only $ 189 million (about $ 215 million) in June, called “compactFi” on the strength of model compression technology. (Since it was founded in 2019, it has raised about $ 250 million.)
And the compact, is a quantum -affected compression algorithm that reduces the size of existing AI models without sacrificing the performance of these models.
Taxkarnch event
San Francisco
|
October 27-29, 2025
He explained, “We have a compression technology that does not have ordinary compression technology that people of computer science or machine learning will do, because we come from quantum physics.” “This is a more subtle and better compression algorithm.”
The company has already released a long list of the compressed version of open source models, especially the popular small models such as Lama 4 Scout or Mr. Small 3.1. And he just launched the compressed version of two new open models of the open. For example, it has also compressed some very large models.
But since it is in the business of making models small, it has focused extra on making the smallest and most powerful model possible.
The company says its two new models are so small that they can bring chat AI’s capabilities to any IOT device and work without internet connection. It calls this family a model zoo with humor as it is named the product based on the size of the animal’s brain.
A model called Super Fly is a compressed version of the open source model of the throat facial Smollm2-135. There are actually 135 million parameters and was designed to use the device. The Super Fly is 94 million parameters, which describe the size of the brain of the bee. “It’s like flying, but a little more smart,” he said.
The Super Fly is designed to train on extremely limited data like a device’s operations. Multipuria has imagined that it has embedded in domestic appliances, which facilitates users with a voice command, such as “Start Quick Wash” for the washing machine. Or users can ask questions that defects. With a little processing power (such as an Urdu), the model can handle the sound interface, as the company shows the Tech Crunch directly into the demo.
The second model is named Chakbarin, and is larger in 3.2 billion parameters, but is far more capable and has the ability to argue. Multiwaris says it is a compressed version of Meta Lama’s 3.1 8B model. Yet it is so small to run on MacBook, no internet connection is needed.
More importantly, Aurus said that Chakbarin actually improves the original in several standard benchmarks, including language skill benchmark MMLU Pro, Mathematics Mathematics Benchmark Mathematics 500 and GSM8, and Normal Academic Benchmark GPQ Hira.
Here are the results of internal tests on the benchmarks. The company did not present the results of the benchmark for the super fly, but multi -virus is also not targeting the super fly in terms of use, which requires reasoning.
It is important to note that Multiwaris is not claiming that its model zoo will defeat the largest sophisticated model on such standards. The zoo performance may not even come to the leader boards. The company says its tech can shrink the size of the model without any performance.
Oras says the company is already interacting with all leading device and equipment makers. “We are talking to Apple,” he said. We are also talking to Samsung, with Sony and HP, obviously. HP came as an investor in the final phase. ” The period was led by the leading European VC firm Bill Hound Capital, which included many others, including HP -Tech Ventures and Toshiba.
Startup machine also offers compression tech for other forms of learning, such as photo identity, and clients like BASF, Eli, Moody, Bush, and in six years. Other
In addition to selling large device manufacturers directly to their models, Multiwaris offers it Compressed model Anybody who can use the API on AWS, who can use the developer, on a lower token fee than most rivals.
We are always looking for evolution, and by providing some insights about teching and your vision and opinions in your coverage and events, you can help us! Fill This survey To tell us how we are doing and have the opportunity to win the prize in return!