DataCentreNews Ireland - Specialist news for cloud & data centre decision-makers
Ireland
OVHcloud unveils Olala mini, its first in-house LLM

OVHcloud unveils Olala mini, its first in-house LLM

Wed, 18th Nov 2026 (Today)
Mara Sugue
MARA SUGUE News Editor

OVHcloud has introduced Olala mini 7A1B, the first model in its new Olala family of large language models and the company's first in-house LLM family.

The compact model is built on what OVHcloud described as a novel architecture and is part of a broader push by European groups to build domestic artificial intelligence technology. Trained from scratch, it is presented as evidence that European businesses can develop agentic LLMs without relying on external labs.

Olala mini 7A1B uses a 7-billion-parameter architecture with one billion active parameters spread across 256 experts in each layer. OVHcloud said it combines Differential-TPA and Mamba-3 through a hybrid attention design in BF16 format.

Only part of the model is used for each request. OVHcloud said this reduces memory demands and allows the model to run on a GPU with 8 GB of memory, delivering 160 tokens per second on an RTX A4500, while a standard processor can reach 50 tokens per second without a GPU.

The context window currently reaches 32,000 tokens. OVHcloud plans to release the model as open source on Hugging Face under the Apache 2.0 licence.

Public backing

OVHcloud trained Olala mini on three European public supercomputers: Jupiter, MeluXina and Jean Zay. Those systems were equipped with GH200, A100 and H100 GPUs.

It valued the public support at EUR €7 million in computing hours, saying the backing helped it build its artificial intelligence team and contributed to the creation of its Agentic Lab, OVHai LLM.

The release includes the model weights, a pre-trained base model, a technical report and AI Act compliance documentation. That may matter to customers and public-sector users in Europe as companies prepare for tighter scrutiny of AI systems and their governance.

Early customer use

Qwant has already adopted the model for its flash answer feature, according to OVHcloud. The deployment produced answers 20% faster and cut running costs to one-third of the previous model's level, it said.

The Qwant example gives OVHcloud an early production reference beyond internal testing. It also suggests the company is trying to compete on running costs as well as model design at a time when many AI deployments remain expensive to maintain.

Broader plan

OVHcloud expects version 1.1 by the end of the year, with a larger context window of 128,000 tokens and improved mathematical reasoning. Version 2.0 would then focus on three areas: agentic AI for businesses, cloud DevOps and cybersecurity.

In business software, the Olala line is intended over time to support OVHai Workspace, including an encrypted version that would require models able to run locally on a workstation. In cloud operations, OVHcloud plans variants tailored to developers, architects and operations teams. In cybersecurity, the focus would cover both defensive and offensive uses.

OVHcloud added that Olala v2.0 will exceed 100 billion parameters. That would be a significant step up from the compact design of the initial release and move the project into a much larger model category.

The announcement comes as European technology companies and policymakers place greater emphasis on sovereign AI infrastructure, local model development and reducing dependence on a small number of US-based AI providers. Access to public supercomputing resources has become one of the main tools available to regional developers seeking to narrow that gap.

For OVHcloud, the move also extends its role beyond cloud infrastructure into model creation. The group is Europe's largest cloud provider and operates 46 data centres worldwide, serving 1.7 million customers in more than 140 countries, according to the company.

"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," said Octave Klaba, Chief Executive Officer, OVHcloud.