OVHcloud unveils Olala, a small model designed for efficiency

OVHcloud introduces Olala, its first family of language models, starting with Olala mini 7A1B. The model, trained on European public supercomputers, targets fast and computationally efficient uses.

On November 17, 2026, OVHcloud announced Olala, its first family of language models, beginning with Olala mini 7A1B. This small model, trained on European public supercomputers, can run with 8 GB of GPU memory according to the company and is to be released open source on Hugging Face under the Apache 2.0 license. The challenge is twofold: reduce execution costs and maintain control over a key AI component, as companies seek models less dependent on major American labs.

A small model, but not a small ambition

Olala mini features an architecture of 7 billion parameters, with 1 billion activated per request. In practice, this expert structure limits memory load and aims for faster responses. OVHcloud reports achieving 160 tokens per second on an RTX A4500 and 50 tokens per second on a standard CPU without GPU. The model also has a context window of 32,000 tokens, meaning the amount of text it can process at once.

The technical choice combines Differential-TPA and Mamba-3, with a BF16 format. On paper, this looks like very serious optimization. In practice, the company mainly aims to fit AI into environments where hardware costs must not skyrocket.

Le pari européen d’OVHcloud

The training was done from scratch on three European public supercomputers: Jupiter, MeluXina, and Jean Zay. OVHcloud also mentions GH200, A100, and H100 GPUs in this infrastructure. The company estimates the public support mobilized at the equivalent of 7 million euros in compute hours, a contribution it considers decisive to launch the project.

This European base is not just a marketing argument. It also anchors OVHcloud’s development in a logic of technical sovereignty, as the company says it wants to build its own models instead of relying on the goodwill of other players to obtain open source LLMs. The message is clear, without much detour.

Qwant sert de premier test grandeur nature

Qwant already uses Olala mini for its flash response function. OVHcloud claims that responses are 20% faster and that operating costs have been cut by three compared to the previously used model. This first deployment matters more than the technical sheet, as it signals the real economy of the model.

The company also plans to publish, along with the model, the weights, a pre-trained base model, and compliance documentation related to the AI Act, the European legislation on artificial intelligence. For client companies, this part will count almost as much as raw performance: a model can be fast, but if it is impractical on the compliance side, it quickly loses its appeal.

La feuille de route monte rapidement en puissance

OVHcloud plans a version 1.1 by the end of 2026, with a context window increased to 128,000 tokens and better results in mathematical reasoning. A version 2.0 is announced above 100 billion parameters, with specialized models for business agentic AI, cloud DevOps operations, and cybersecurity, including both defensive and offensive according to the roadmap. The project thus goes beyond the simple launch of a compact model.

The company is also preparing internal uses, notably OVHai Workspace, whose encrypted version must run locally on workstations, as well as tools for developers, architects, and operations teams. Octave Klaba sums up the idea as follows: « We cannot simply rely on the generosity of certain labs to share their LLMs as open source in the years to come, » according to the quote cited by OVHcloud. He adds that training costs have been divided by 20 in three years, a central argument to make this type of development sustainable.

Key points

  • Olala mini was announced on November 17, 2026.
  • Qwant reports a 20% speed gain on flash responses.
  • Operating costs were reportedly cut by three at Qwant.
  • OVHcloud aims for more than 100 billion parameters in version 2.0.
  • Octave Klaba leads the OVHai LLM strategy.
  • The model was trained on Jupiter, MeluXina, and Jean Zay.

By the numbers

  • 7 billion — parameters of Olala mini 7A1B, according to OVHcloud.
  • 8 GB — minimum GPU memory announced for execution.
  • 32,000 tokens — current context window of the model.
  • 160 tokens/s — throughput reported on an RTX A4500.
  • 7 million euros — equivalent of public compute hours mobilized.

To read