Thursday, November 19, 2026

Top 5 This Week

Related Posts

Olala mini: OVHcloud launches its first language model, open and usable locally


So far, OVHcloud rented the artificial intelligence infrastructure: servers and graphics processors on which other companies trained and ran their models. The Roubais group now moves to the other side. THE November 17, 2026he revealed Olala mini 7A1Bthe first representative of his own family of major language modelstrained from scratch and published in open access. “What we want to show is that Europe has its place in all layers of artificial intelligence,” summarizes the company, quoted by the Moderator’s Blog.

A small model designed to go anywhere

Olala mini chooses sobriety rather than raw power. It is a model to expert mix of 7 billion parametersof which only one billion is active at each stage, hence its name, 7A1B. Each layer has 256 experts, specialized subnetworks of which only a part is requested for each piece of text.

The architecture is hybrid: it alternates layers of attention of type Differential-TPAwhich identify useful passages of the text to produce the answer, and layers Mamba-3which keep a compact representation of what has already been read without re-examining everything. Objective: to reduce the memory and calculation required.

Characteristic Olala mini 7A1B
Settings 7 billion, including 1 billion active
Architecture Mix of experts, 256 experts per layer, hybrid attention Differential-TPA and Mamba-3
Context window 32,000 tokens
Minimum equipment announced Graphics card with 8 GB of memory
Announced speed 160 tokens per second on an RTX A4500, 50 on processor alone
License Apache 2.0, commercial use permitted
Publication November 19, 2026 on Hugging Face

Concretely, the model can work locallyon a user’s computer or a modest server, without sending the data to a remote service. An important argument for companies that handle confidential documents. The intended uses are those of small models: summarizing texts, answering questions on internal documents, classifying messages or powering functions integrated into software, such as flash responses from a search engine. However, for complex reasoning or long documents, a model of this size remains limited, especially since its context window is still modest.

Trained on European public supercomputers

OVHcloud did not start from an existing model. Olala mini was trained from scratch on three public supercomputers: Jupiterin Germany, MeluXinain Luxembourg, and Jean ZayIn France. This public support represents the equivalent of 7 million euros in computing hours, according to the company.

The model is the result of the laboratory OVHai LLM. According to Le Filia, this is the new name of Dragon LLM, formerly Lingua Custodia, a Parisian SME founded in 2011 and specializing in financial translation, of which OVHcloud announced the acquisition in March 2026.

For Octave Klaba, president of OVHcloud, the economic equation has changed: according to him, the cost of training “has been divided by 20 in 3 years”. He adds, quoted by Alliancy: “We cannot just count on the generosity of certain labs to share their LLM models in open source in the years to come. »

A first client: Qwant

Olala mini is not just a technical demonstration. The French search engine Qwant already uses it for its flash answersthe summaries generated at the top of the search results. According to OVHcloud, these responses are 20% faster and cost three times cheaper than with the model used previously, which has not been named.

The model will be published with its weighther pre-trained base modelA technical report and one EU AI Regulation compliance documentation. This last point directly targets European companies, required to document the models they deploy.

Olala mini: promises and unknowns

What is announced

  • An open model under Apache 2.0 license, usable commercially
  • Local operation, with 8 GB of graphics or processor memory
  • Zero training on European supercomputers
  • Documentation of compliance with the European regulation on AI

What remains to be demonstrated

  • The quality of the answers, lack of published score
  • The exact conditions of the announced speed measurements
  • The outfit compared to the small models of Mistral AI, Qwen or Gemma
  • Adoption beyond Qwant, the only client cited

What still needs to be checked

The announcement presents a major gap: no performance score has not been published. OVHcloud communicates throughputs, not results on benchmark tests. Impossible, at this stage, to compare Olala mini to small open models of the same size, such as those of the Qwen, Gemma family or the compact models of Mistral AIavailable in 3, 8 and 14 billion parameters.

Filia also notes an apparent inconsistency: in BF16 precision, seven billion parameters occupy around 14 GB of memory, almost double the 8 GB announced. The 8 GB figure suggests a compressed version of the model, without knowing whether the announced speeds were measured with it. The processor used for the measurement of 50 tokens per second was also not specified.

However, a European point of comparison exists: according to Le Filia, the Italian Translated published Lara Mini Base on November 12, an expert mixture of 7.2 billion parameters including 0.6 billion active, accompanied by its scores, but under a license which prohibits commercial use. Olala mini will have the advantage of licensing, provided the quality follows.

The OVHcloud roadmap

Olala mini is just the first step. OVHcloud detailed the rest:

  • version 1.1expected by the end of 2026: context window increased from 32,000 to 128,000 tokens and better mathematical reasoning;
  • version 2.0: more than 100 billion parametersavailable in specialized models in three areas, the agents for businessesTHE Cloud DevOps for developers and operations teams, and cybersecuritydefensive and offensive according to Clubic;
  • integration into OVHai Workspacethe company’s working environment, a future encrypted version of which will be based on models that can be executed directly on the workstation.

According to Reuters, cited by ICT Journal, OVHcloud plans to invest between 150 and 200 million euros in its own family of models.

A second French actor facing Mistral AI

With Olala, OVHcloud becomes the second French player to develop its own language models alongside Mistral AI, but with a different economic model. The profitable group says it does not need to raise funds to finance its AI strategy, whereas Mistral AI relies on successive fundraising. The resources remain incommensurable: OVHcloud employs just under 100 people on AI, compared to more than 1,000 at Mistral AI according to AFP, cited by Le Filia.

For French companies and administrations concerned about digital sovereignty, the arrival of an open model, trained in Europe and offered by a European host, broadens the choice. The main thing remains: the quality of the answers. The publication of the weights on November 19, during the OVHcloud AI Summit, will finally allow researchers and developers to measure what Olala mini is really worth compared to the competition.

Frequently asked questions

What is Olala mini from OVHcloud?

Olala mini 7A1B is OVHcloud’s first large language model, unveiled on November 17, 2026. It is an expert mixture model of 7 billion parameters, including one billion active at a time, trained from scratch on European public supercomputers and published in open access under the Apache 2.0 license.

Can you use Olala mini on your computer?

Yes, according to OVHcloud. The model works with a graphics card with 8 GB of memory and can even run on a single processor, at around 50 tokens per second. His weights are published on November 19, 2026 on Hugging Face. This 8 GB figure probably assumes a compressed version of the model.

Is Olala mini better than Mistral AI models?

We don’t know yet. OVHcloud has not published any performance scores, only generation speeds. It is therefore impossible to compare it to the compact models of Mistral AI, Qwen or Gemma. The first independent tests should follow the publication of the weights on November 19, 2026.

Is Olala mini free for commercial use?

Yes. The model is released under the Apache 2.0 License, which permits reuse and modification, including for commercial purposes. OVHcloud also provides the pre-trained base model, a technical report and documentation of compliance with the European AI regulation, useful to companies deploying it.

Sources

References consulted for the writing of this article. The addresses are provided for information purposes only and are not links.

  1. Moderator’s Blog, Frenchman OVHcloud launches into LLM with Olala, a small open model, November 17, 2026www.blogdumoderateur.com/ovhcloud-llm-olala-petit-modele-ouvert
  2. Clubic, OVHcloud launches Olala, its first in-house AI model trained on European supercomputers, November 17, 2026www.clubic.com/actualite-639616-ovhcloud-lance-olala-son-premier-modele-d-ia-maison-entraine-sur-des-supercalculateurs-europeens.html
  3. Le Filia, OVHcloud unveils Olala mini, its first LLM: throughput, not scores, November 2026www.lefilia.fr/article/6482241-ovhcloud-devoile-olala-mini-son-premier-llm-des-debits-pas-de-scores
  4. Alliancy, OVHcloud trains Olala mini, its first in-house LLM, and targets sovereign agentic AI, November 2026alliancy.fr/ovhcloud-entraine-olala-mini-son-premier-llm-maison-et-vise-lia-agentique-souveraine-c6b80086-fc53-401b-b984-ca42a10f18fa

Report an error or inaccuracy in this article

Shredder Smith
Shredder Smith
Shredder Smith is the lead curator and digital persona behind topaitools4you.com, an AI directory dedicated to "shredding" through industry hype to identify high-utility software for everyday users. Smith positions himself as a blunt, no-nonsense reviewer who vets thousands of emerging applications to filter out overpriced "wrappers" in favor of tools that offer genuine ROI and practical productivity. The site serves as a watchdog for the AI gold rush, providing categorized rankings and transparent reviews designed to help small businesses and creators navigate the crowded tech landscape without wasting money on low-value tools.

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Popular Articles