Portrait of OVHcloud founder and CEO Octave Klaba in a dark suit and glasses against a dark background
OVHcloud (Asset Library)

AI Europe

OVH launches Olala, an open model trained on MeluXina

OVH's first language model, Olala mini 7A1B, is set to be released under Apache 2.0 and was trained partly on Luxembourg's MeluXina supercomputer.

By Nova 5 min read

Key takeaways

  • OVH announced Olala mini 7A1B on 17.11.2026 and launches it on 19.11.2026 at its AI Summit in Paris.
  • The model has 7 billion parameters with about 1 billion active and is to be published under Apache 2.0.
  • It was trained on Jupiter, MeluXina and Jean Zay with public compute worth the equivalent of 7 million euro.
  • There are no independent benchmarks yet, and the supported languages are not documented.
On this page
  1. What is Olala mini 7A1B?
  2. What does open mean here?
  3. Which supercomputers trained it?
  4. Why does MeluXina matter for Luxembourg?
  5. How does it compare with Mistral?
  6. What comes next?
  7. What should Luxembourg companies do?

OVH has announced Olala, its first family of large language models. The first model, Olala mini 7A1B, is a compact open model that the company says it trained from scratch on three European public supercomputers, including MeluXina in Luxembourg.

According to the press release of 17.11.2026, the official launch is set for the OVHcloud AI Summit in Paris on 19.11.2026. The weights are to be published on Hugging Face under the Apache 2.0 licence. On the morning of 19.11.2026, before the summit keynote, no Olala model was yet listed on OVHcloud's Hugging Face page.

What is Olala mini 7A1B?

The name describes the design. Olala mini has 7 billion parameters in total, of which about 1 billion are active for each request, spread across 256 experts per layer. OVH says it combines two attention techniques, Differential-TPA and Mamba-3, and runs in BF16 format. Because only part of the model works on each request, it needs less memory.

OVH gives these figures:

  • It runs on a GPU with as little as 8 GB of memory.
  • It reaches 160 tokens per second on an Nvidia RTX A4500.
  • Without a GPU, on a standard processor, it reaches 50 tokens per second.
  • The context window is currently 32.000 tokens.

These are the company's own statements. The press release contains no quality benchmarks, and no independent tests had been published by the morning of 19.11.2026. The release also does not say which languages the model handles. Companies should check both once the technical report is out.

What does open mean here?

OVH says Olala mini will be published with its weights, the pre-trained base model, a technical report and documentation on compliance with the EU AI Act. Apache 2.0 is a permissive licence that allows commercial use and modification. In practice, a company could download the model, adapt it and run it on its own hardware, without sending data to an outside service.

The French search engine Qwant already uses the model for its quick answers feature. According to OVH, answers are 20 percent faster and running costs are a third of those of the previous model.

Weekly, Thursday morning, Central European Time.

The Thinking briefing

One email a week: what changed in European business software, which prices moved, and which comparison we published. No sales mail.

Which supercomputers trained it?

OVH names three public machines: Jupiter in Germany, MeluXina in Luxembourg and Jean Zay in France, with Nvidia GH200, A100 and H100 GPUs. The company values this public support at the equivalent of 7 million euro in computing hours and calls it decisive. The release does not break the hours down by machine or name the programme that granted them.

Technician in an OVHcloud data centre aisle between server racks full of coloured network cables
OVHcloud runs its own data centres, but trained Olala mini on public supercomputers. Archive picture from Roubaix | Image: OVHcloud (Asset Library)

The model comes out of OVH's AI lab, now called OVHai LLM. Its roots lie in Dragon LLM, formerly Lingua Custodia, which OVHcloud agreed to buy in March 2026. As Lingua Custodia, the company had won the European Commission's Large AI Grand Challenge, which came with computing time on EuroHPC supercomputers.

Why does MeluXina matter for Luxembourg?

MeluXina is the national supercomputer in Bissen, operated by LuxProvide. A commercial European model trained partly on it shows that the Luxembourg machine is used for more than research projects. Its successor is already planned: MeluXina-AI, an AI-optimised system for the Luxembourg AI Factory with an estimated investment of 112 million euro, shared equally between the EuroHPC Joint Undertaking and the Luxembourg State. We covered that project in our article on MeluXina-AI and the AI Factory and on the EuroHPC call for large-scale AI access.

The MeluXina supercomputer in Bissen, Luxembourg
MeluXina in Bissen was one of the three public supercomputers used to train Olala mini | Image: © SIP / Julien Warnand (gouvernement.lu)

How does it compare with Mistral?

Mistral AI is the obvious reference point, and in Luxembourg the closer partner: the government signed a strategic partnership with Mistral in June 2025, with on-premise solutions for the state. Mistral already sells small open models. Its documentation lists Ministral 3 in 3B, 8B and 14B sizes under Apache 2.0, and Mistral Small 4 also under Apache 2.0, while Mistral Large 4 uses Mistral's own open weight licence.

On paper, Olala mini competes in the same licence class as the Ministral models. What is different is the route: OVH trained its own architecture rather than buying in a model, and it is a cloud provider that also hosts other labs' models. Which model works better for a given task cannot be said from the announcements. Our analysis of Mistral and European software explains what questions to ask a supplier before relying on any model.

What comes next?

OVH plans version 1.1 by the end of 2026, with a context window of 128.000 tokens and better maths reasoning. Version 2.0 is to exceed 100 billion parameters, with specialised models for agentic AI in companies, cloud DevOps and cybersecurity. In time, Olala is to power OVHai Workspace, including an encrypted version that runs models locally on the workstation.

CEO Octave Klaba framed it as a matter of independence: "We cannot simply rely on the generosity of certain labs to share their LLMs as open source in the years to come," he said. He added that the cost of training has been divided by 20 in three years.

What should Luxembourg companies do?

Nothing is final yet. Companies that want a small model they can run on their own servers should wait for the files and the technical report, then test Olala mini on their own tasks against alternatives such as Ministral 3. The questions are the usual ones: which languages it handles well, what the AI Act documentation contains, and who maintains the model. The wider debate on European compute and cloud capacity is covered in our article on the Cloud and AI Development Act.

Sources

Every price and fact above was checked against these pages on the date shown.

  1. [1] OVH unveils Olala mini, its first open source LLM (press release) accessed 2026-11-19
  2. [2] OVH dévoile Olala mini (communiqué de presse) accessed 2026-11-19
  3. [3] OVHcloud AI Summit 2026 accessed 2026-11-19
  4. [4] OVHcloud on Hugging Face accessed 2026-11-19
  5. [5] OVHcloud to acquire Dragon LLM and launch its AI lab accessed 2026-11-19
  6. [6] OVHcloud s'offre Dragon LLM accessed 2026-11-19
  7. [7] Dragon LLM: Europe's frugal AI contender accessed 2026-11-19
  8. [8] L'entreprise française OVHcloud dévoile son premier modèle d'IA baptisé Olala (AFP) accessed 2026-11-19
  9. [9] L'entreprise roubaisienne OVH se lance dans l'IA avec son propre modèle accessed 2026-11-19
  10. [10] OVHcloud launches European-trained open source AI model accessed 2026-11-19
  11. [11] OVHcloud dévoile un LLM compact destiné aux usages professionnels accessed 2026-11-19
  12. [12] OVHcloud moves up AI value chain with launch of first in-house LLM accessed 2026-11-19
  13. [13] MeluXina-AI accessed 2026-11-19
  14. [14] Partenariat stratégique entre le Luxembourg et Mistral AI accessed 2026-11-19
  15. [15] Mistral AI models overview accessed 2026-11-19
  16. [16] OVHcloud AI Endpoints accessed 2026-11-19
Topics OVHcloudOlalaopen source AIMeluXinaEuroHPCMistral AILuxembourg