OVHcloud has launched Olala, its first family of large language models. The first release, Olala mini 7A1B, is the company's first open-source LLM.
The model was trained on three European public supercomputers: Jupiter, MeluXina and Jean Zay.
The launch marks a move by the French cloud provider to build and publish its own AI models, as European groups weigh how far they depend on US-based infrastructure and model developers. OVHcloud said Olala mini 7A1B was trained entirely on European systems and presented the release as evidence that regional companies can build models from scratch.
Olala mini 7A1B is a compact model built on a 7-billion-parameter architecture, with one billion active parameters spread across 256 experts per layer. The company said the design uses only a fraction of the model for each request, reducing memory use and helping it run on more modest hardware.
According to OVHcloud, the model can run on a GPU with 8GB of memory or on a standard processor without a GPU. It reaches 160 tokens per second on an RTX A4500 and 50 tokens per second on a CPU, with a context window of 32,000 tokens.
The release includes the model weights, pre-trained base model, technical report and AI Act compliance documentation. OVHcloud said Olala mini will be made available on Hugging Face under an Apache 2.0 licence.
Early use
One early user is French search engine Qwant, which has adopted the model for its flash answer feature. OVHcloud said Qwant recorded answers that were 20% faster than with its previous model, while running costs were cut by a factor of three.
The company argues that lower hardware requirements could widen the range of organisations able to deploy language models locally or on existing infrastructure. That is likely to matter for businesses trying to control spending as AI workloads grow, and for users with data-handling requirements that limit reliance on external platforms.
Training costs
Octave Klaba, Chief Executive Officer of OVHcloud, said: "AI is strategic for OVH. 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. 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."
The launch also broadens OVHcloud's role in AI beyond infrastructure and managed services. Best known as a European cloud provider with 46 data centres and 1.7 million customers in more than 140 countries, the company is adding a software layer to its AI offering.
Its emphasis on open-source licensing may also strengthen its appeal to developers and businesses that want more visibility into how models are built and deployed. Publishing the base model, weights and documentation gives users the option to inspect, adapt or run the model in their own environments.
OVHcloud said later versions of the Olala line will focus on business agents, cloud operations and cybersecurity. It also said a future version of the family will exceed 100 billion parameters.