OVHcloud launches open AI model trained entirely on European supercomputers
Olala mini 7A1B comes with an Apache 2.0 licensing commitment and modest hardware requirements. OVHcloud says early customer Qwant cut costs and sped up answers.
Listen to this story
The audio brief
Story brief
3 key pointsOVHcloud’s first Olala release gives the cloud provider an in-house model-building capability amid concern that other labs may become less willing to share open models. The 7-billion-parameter, 1-billion-active model was trained on three European public supercomputers and comes with weights, a base model and AI Act compliance documentation under Apache 2.0. Its sparse design is intended to make GPU-free or 8GB-GPU deployment practical, though the published speed figures are company claims.
- 01
OVHcloud reports 50 tokens per second on a standard CPU and 160 on an Nvidia RTX A4500 GPU; these are company figures, not independent comparisons.
- 02
Qwant uses Olala mini for its flash answers and reports responses 20% faster and running costs three times lower than with its previous model.
- 03
The current context window is 32,000 tokens; OVHcloud plans to raise it to 128,000 before the end of 2026.
OVHcloud has launched its own language-model family, starting with a compact model it says was trained entirely on European supercomputers. The first release, Olala mini 7A1B, is designed to run without a graphics processor. The company is pairing it with an Apache 2.0 licensing commitment and plans to distribute it through Hugging Face.
For the cloud provider, the release is about maintaining the ability to build models, not just offering access to them. CEO Octave Klaba said OVHcloud needs that capability for its internal work and its customers. He framed the investment partly as protection against other laboratories becoming less willing to share models openly.
We cannot simply rely on the generosity of certain labs to share their LLMs as open source in the years to come.
Octave Klaba, OVHcloud CEO
Olala mini 7A1B has seven billion parameters—the learned values that make up a model—but only one billion are active. Its architecture has 256 experts per layer. OVHcloud says it uses only a fraction of the model for each request, reducing memory demands and helping it run on more modest hardware.
The hardware pitch gives businesses two routes. According to OVHcloud, the model can run on a graphics processor with 8GB of memory, or on a standard central processor without a GPU. That makes a dedicated graphics card optional for running this first member of the family.
OVHcloud puts output speed at 50 tokens per second on a standard CPU and 160 on an Nvidia RTX A4500 GPU. Tokens are the chunks of text a model processes or produces. These are the company's performance figures, rather than an independent comparison of the two deployment options.
The current context window is 32,000 tokens: the amount of text the model can work with at once. OVHcloud expects a subsequent version to raise that to 128,000 before the end of 2026. The larger window is a planned upgrade, not a capability of the current release.
The training infrastructure is central to OVHcloud's positioning. It says Olala mini was trained exclusively on Jupiter, MeluXina and Jean Zay, three European public supercomputers. Alongside the trained model, the release includes supporting materials that extend beyond a downloadable set of weights:
- Model weights and a pre-trained base model.
- A technical report describing the model.
- AI Act compliance documentation.
Publishing the base model, weights and documentation gives developers options to inspect, adapt or run it in their own environments. OVHcloud says distribution will be through Hugging Face under Apache 2.0. That commitment concerns the release package and its licensing, while the European-training claim concerns the infrastructure used to build it.
French search engine Qwant has adopted Olala mini for its flash answer feature. OVHcloud says answers were 20% faster than with Qwant's previous model, while running costs fell by a factor of three. Those gains describe this particular customer deployment; they are not a general performance ranking against other language models.
The family is not intended to stay confined to this compact starting point. OVHcloud says later versions will target business agents, cloud operations and cybersecurity. It also plans a future member exceeding 100 billion parameters—placing a much larger model on the roadmap, rather than in this initial release.
Sources
- itpro.comOVHcloud launches European-trained open source AI model
- itbrief.co.ukOVHcloud launches first open-source AI model family
Reader comments
Newest comments first. Replies stay oldest first.