A Basque startup of AI has just lifted 189 million euros with a great idea: compress the AI

Before We compressed files with Zip. Now what we begin to need is to compress the AI ​​to make it more small and efficient. That is just the idea that the founders of Multivrse Computing had, a Spanish startup which is becoming the new jewel of the crown of our AI industry. Its founders, (in the image, from left to right, Román Orús, Enrique Lizaso Olmos and Samuel Mugel) and Alfonso Rubio have much to celebrate.

Investment Round. Multivize Computing He has just closed an investment round of 189 million euros (215 million dollars). The round (series B) has been led by Bullhound Capital, but it has also participated HP Tech Ventures, Sett, ForgePoint Capital International, CDP Venture Capital, Santander Climate VC, Quantonation, Toshiba and Euskadi Risk Capital of Euskadi – Spri Group. Last March the company received An investment of 67 million euros by the Government of Spain.

The inference AI by flag. Although the current prominence usually takes it the great technological ones that invest billions of dollars in data centers to train Great language models (LLM)there is more and more focus on the other part: the one we use users when asking things to Chatgpt, for example. It is the so -called AI inference, and the estimate is that in 2025 the value of that industry reaches 106,000 million dollars. In Multivrse Computing they want a good piece of that cake, and to achieve this, its great trick is a unique technology.

Compactifai. This is the name of The compression technology of AI models developed by multivance computing. What this allows is to convert very large models – which costs a lot to “execute” – in much smaller and efficient models, which allows them to make them more manageable and save many resources (and time) during inference.

How to compress an AI model. Román Orús, scientific director of the company, led A study May 2024 in which they precisely explained the concept of “tensioning networks” of quantum inspiration and that allow compressing these models. Its operation is based on decomposing the matrices of pesos from the neural networks “truncating them” and retaining only the largest and most relevant values. In essence the concept focuses on discarding the less relevant information of the model to be left alone with the most relevant.

But that does not make the model less accurate? In fact, but the degree of truncation can be controlled so that there is a good balance and commitment between compression and loss of precision. Even by compressing these models, in Multivars Computing they say The fall of the models It is only 2 to 3%.

Same yield in a size 95% lower. To mitigate that precision fall, this system includes a rapid resentment phase called “healing” that can be repeated several times to achieve even closer accuracy to the original version. In the end, they affirm in the company, they can compress up to 95% a model of the performance.

It lowers the use of AI. According to Your dataa model as it calls 3.1 405b has an operational cost of about $ 390,000 if we want to run it at home (13 GPUS H100, 9100 W of consumption), but thanks to Compactifai it is possible to reduce that cost to 60,000 dollars (2 GPUS H100, 1,400 W).

One more “thin”. The “Slim” models provided by the company – Derivatives of Llama 3.3 70b or Call 4 scout– They are compressed versions that theoretically do not lose precision. They can be executed through the AWS platform or by licenses that also allow us to use it on-premisethat is, in local/own infrastructure. According to their metrics, these models are between 4 and 12 times faster than their non -compressed versions, which translates into an inference cost that is between 50% and 80% lower.

Image | Multivize Computing

In Xataka | Spain is finally