OVHcloud launches European-trained open source AI model

IT Pro

Source published: Site published:

Original article · فارسی

The new small AI language model can run on a regular desktop computer's CPU. OVHcloud plans to increase the amount of text it can handle at once to 128,000 tokens.

Adam Shepherd's avatar
By Adam Shepherd
Published In News 3 min read
OVHcloud logo and branding pictured on a company data center in Roubaix, northern France.
(Image credit: Getty Images)
Newsletter

OVHcloud has launched a new group of open source AI models that were trained only using computer systems in Europe.

The new group of large language models (LLMs) is called Olala. The first model released is the small Olala mini 7A1B. It was trained only on European supercomputers, including Jupiter, MeluXina, and Jean Zay.

OVHcloud says it developed the model partly because it expects large technology companies may stop sharing their open source models in the future.

“AI is strategic for OVH,” said CEO Octave Klaba. “As well as offering AI services to our customers, we must be able to create models for our own internal needs and for those of our customers.”

“We cannot simply rely on the generosity of certain labs to share their LLMs as open source in the years to come,” Klaba added. “Know-how in model training has become widely accessible, and the cost of training has been divided by 20 in three years. It is now possible to create LLMs while making this activity profitable.”

How OVHcloud’s new model works

The model uses a new design intended to make it fast while using less memory. OVHcloud says Olala mini 7A1B can run on a standard computer processor (CPU) at 50 tokens per second. It can also run on an Nvidia RTX A4500 graphics card (GPU) at 160 tokens per second.

The model has 7 billion parameters and can currently handle a context of 32,000 tokens at a time. OVHcloud expects the next version to increase this limit to 128,000 tokens before the end of the year.

Olala mini 7A1B will be released as open source under an Apache 2.0 license through Hugging Face. It will come with the model's weights, technical details, and documents showing how it follows the EU AI Act.

The French search engine Qwant is already using Olala mini 7A1B. Reports say it gives answers 20% faster than the model it replaced, while its running costs are three times lower.

OVHcloud plans to keep developing AI models for specific business tasks. These include AI agents running on local devices, cloud DevOps tasks, and cybersecurity.

Digital independence is under scrutiny

The announcement comes as more people question the relationship between European customers and technology companies in the US. Many organizations and governments in the EU are trying to reduce their dependence on American technology.

Gartner says that complete digital independence may not be realistic. However, by 2030, every country is expected to see reducing dependence on foreign technology as an important national security issue.

AI has become a major part of this debate. The US government has shown that it is willing to use access to AI models and computer chips as economic tools to support its own interests.

In response, the EU has invested in building large AI computing centers across Europe, known as AI gigafactories. Meanwhile, the UK government has launched a £500 million investment fund to support British AI startups.