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

One of the most prominent AI startups in Europe has released two AI models that are so small, they have named them after the brain of a chicken and the brains of a fly.

Multiversum Computing Claims these are the world’s smallest models that still perform well and in one case can handle chat, speech and even reasoning.

These new small models are intended to be embedded in Internet of Things devices, and are performed locally on smartphones, tablets and PCs.

“We can compress the model in such a way that they can fit on devices,” founder Román Orús told WAN. “You can perform them on buildings, directly on your iPhone or on your Apple Watch.”

As we have reported earlier, Multiversum Computing is a Buzzy European AI Startup with head office in Donostia, Spain, with around 100 employees in offices worldwide. It was co -founded by a top European professor of quantum computers and physics, Román Orús; Quantum Computing expert Samuel Mugel; en enrique Lizaso Olmos, the former deputy CEO of Unnim Banc.

In June it raised € 189 million (around $ 215 million) on the basis of a model compression technology that calls the ‘compactifai’. (Because it was founded in 2019, it collected around $ 250 million, said Orús.)

Compactifai is a quantum-inspired compression algorithm that reduces the size of existing AI models without sacrificing the performance of those models, Orús said.

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

The company has already released a long list of compressed versions of Open Source models, especially popular small models such as Lama 4 Scout or Mistral Small 3.1. And it just launched compressed versions of the two new open models from OpenAi. It has also compressed a number of very large models – for example, it offers a Deepseek R1 smart.

But because it is busy making models smaller, it has extra attention aimed at making the smallest possible but powerful models possible.

The two new models are so small that they can bring chat AI options to just about every IoT device and working without an internet connection, says the company. It calls this family humorous the model Zoo because it calls the products based on brain sizes for animals.

A model that calls the Superfly is a compressed version of the open source model Smollm2-135 from Hugging Face. The original has 135 million parameters and has been developed for use on the device. Superfly is 94 million parameters, which compares Orús with the size of the brain of a fly. “This is like having a fly, but a little smarter,” he said.

Superfly is designed to be trained on very limited data, such as the operations of a device. Multiversum sees it embedded in household appliances, so that users can operate them with speech assignments such as “Start Quick Wash” for a washing machine. Or users can ask for solving problems. With a little processing power (such as an Arduino), the model can handle a speech interface, as the company showed in a live demo from WAN.

The other model is called Chickbrain and is greater at 3.2 billion parameters, but is also much more capable and has opportunities. It is a compressed version of the Lama 3.1 8B model from Meta, says Multiversum. Yet it is small enough to run on a MacBook, does not require an internet connection.

What is even more important is that Orús said that Chickbrain performs the original in various standard benchmarks somewhat better, including the Language Proficiency MMLU-Pro, mathematical skills Benchmarks Math 500 and GSM8K and the General Knowledge Benchmark GPQa Diamond.

Here are the results of the internal tests of Multiversum from Chickbrain on the benchmarks. The company did not offer a benchmark results for Superfly, but Multiversum also does not focus on superfly in use cases that require reasoning.

It is important to note that Multiversum does not claim that the model will defeat the greatest ultraming of models on such benchmarks. Implementation may not even land on the leaderboards. The point is that the technology can shrink the model size without a performance hit, the company says.

Orús says that the company is already in conversation with all leading equipment and equipment makers. “We talk to Apple. We talk to Samsung, also with Sony and HP, of course. HP came as an investor in the final round,” he said. De Ronde was led by the well-known European VC company Bullhound Capital, with the participation of many others, including HP Tech Ventures and Toshiba.

The startup also offers compression technology for other forms of machine learning, such as image recognition, and has obtained customers in six years such as Basf, Ally, Moody’s, Bosch, and Others.

In addition to the sale of its models directly to large manufacturers of devices, Multiversum offers its compressed models Via an API on AWS hosted that any developer can use, often at lower token costs than competitors.

We always want to evolve, and by giving some insight into your perspective and feedback in WAN and our coverage and events, you can help us! Fill up This investigation To let us know how we do it and get the chance to win a prize in exchange for that!