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 a chicken brain and a flight brain.

Multivers computing claims that they are the smallest models in the world that can still handle high performance and manage chat, speech and even logic in a case.

These new minor models are for the purpose of running locally on smartphones, tablets and PCs as well as embedded on internet -things devices.

“We can compress the model so much that they can fit on the device,” Ors told TechCrunch. “You can run these directly to your iPhone or your Apple Watch”

As we have already said, Multivars Computing is a Buzzy European AI Startup in Doncistia, Spain, has about 100 employees in the European AI Startup, global offices. It is co-founded by Roman Orris, a top European Professor in Quantum Computer and Physics; Quantum computing expert Samuel Mugel; And the former Deputy CEO of Enrique Lizaso Olmos Anim Bank.

It barely raised 9 189 million (about $ 215 million) A model in June is called “compactify” about the power of comprehension technology. (Since it was founded in 2019, it has collected about $ 250 million, Ors says.)

Compactify is a quantum-inspired algorithm that reduces the size of the existing AI models without leaving the performance of these models, Oras.

“We have a briefing technology that is not the technology of general contraction that people in computer science or machine learning are from quantum physics,” he describes. “It’s more subtle and more refined abbreviation algorithm.”

The company has already published a long list of compressed versions of open-source models, like especially popular small models Call 4 Scouts Or Mistral Small 3.1The And it has just launched compressed versions OpeneThe It also compressed some very large models – it provides a DEPSEC and 1 slim for example.

However, since it is in the business to make the models smaller, it is the smallest but focusing on the most powerful models possible.

Its two new models are so small that they can bring chat AI power to an IoT device and work without internet connection, the company says. It is ridiculously called this family a model zoo because it is naming the products based on the brain size of the animal.

A model it calls Superfly is a compressed version of the open source model Smallm 2 135 of the hug, The origin has 135M parameter and it was made for on-division use. Superfly is a 94 m parameter, which is compared to the brain size of a flight of the Ors. “It’s like a fly, but somewhat more cute,” he said.

Superfly is designed to be trained in a very limited data like a device’s activity. Multivars imagine it embedded on home applications, allowing users to handle them with voice commands like “Quick Wash” for a washing machine. Or users can ask the problem to solve the problem. With a little processing power (like an Arduino), the model can handle a voice interface, as the company showed a live demo on TechCrunch.

The other model is named after Murbren, and it is large in 3.2 billion parameters, but it is more capable and has the logic of its logic. This is a compressed version of the Meta Lama 3.1 8B model, Multivars says. Nevertheless it is small enough to run a MacBook, no internet connection is needed.

More importantly, Ors said that Murbren was actually the Language-Skill Benchmark MMLU-Pro, Math Skill Benchmarks Math 500 and GSM 8K, and General Knowmmark JPQA Diamond, a number of standard benches.

Here are the results of the internal examination of the multiverse in the benchmarks. The company did not provide benchmark results for superfly, but the use of multiverse does not notice the superfly that requires logic.

It is important to note that the multiverse does not claim that its model zoo will defeat the largest sophisticated models in this national standard. Zoo performances cannot even land on leadersboards. The original statement is that its technology can shrink the model size without a performance hit, the company says.

Ors says the company is already in talks with all the top devices and appliance makers. “We’re talking to Apple. We are clearly talking to Samsung, Sony and HP. HP came as an investor in the last round,” he said. The round was led by the participation of well -known European VC Farm Bullhound Capital, HP Tech Ventures and Toshiba.

Startup provides briefing technology for such image recognition for other types of machine learning and has received clients like BASF, Allie, Moody, Bosh and and the likes in six years Others

In addition to selling its models directly to the main device manufacturers, Multivars offers it Compressed model An API is hosted to AWS, which can use a developer, often less token than competitors.