Olala, OVHcloud’s bet to enter language models

OVHcloud unveiled Olala mini, its first family of language models, trained in Europe and released as open source. Without public benchmarks, the group currently emphasizes its technical promises and customer base.

On November 17, 2026, OVHcloud presented Olala mini, its first family of language models, with an official launch announced for November 19 at the OVHcloud AI Summit in Paris. This move places the French cloud specialist in the generative AI arena, previously dominated in France by Mistral AI. The group wants to show it no longer just hosts others’ models. One important detail remains: no public benchmark yet allows judging Olala mini’s real performance.

A compact model, designed to run far from war machines

The first model in this family, Olala mini 7A1B, is a small open-source model designed for relatively light uses. OVHcloud announces 7 billion parameters in total, with 1 billion activated per request, featuring a MoE architecture, meaning a mixture of experts system where only certain subnetworks work at a given moment. The company also aims for execution starting at 8 GB of graphics memory, with 160 tokens per second on an RTX A4500 and 50 tokens per second on a simpler processor. No need for a supercomputer under the desk. That’s the idea.

The context window reaches 32,000 tokens, with an extension to 128,000 tokens announced in version 1.1 expected by the end of the year [to be confirmed]. The model is published under the Apache 2.0 license on Hugging Face, with weights, the pretrained base model, a technical report, and AI Act compliance documentation. OVHcloud also specifies that training was conducted from scratch on European supercomputers.

Why OVHcloud is joining the game now

The timeline has accelerated. In June, Octave Klaba had already mentioned the idea of training its own models, initially planned for early 2027. The group ultimately moved its announcement up to November 2026. Behind this change of pace, the stake is clear: OVHcloud has 1.7 million customers and 46 datacenters, an asset few generative AI players can claim.

This commercial base can become a very concrete distribution channel for Olala. Like Google Cloud with Gemini, OVHcloud can push its own models to companies and sites already hosted with it. In a context where AI sovereignty is back in every discussion, the argument carries weight. And it’s better to arrive with your own building blocks when you already sell the infrastructure.

Uses already tested, but still many gray areas

For now, OVHcloud publishes no benchmarks, making comparison with Google’s Gemma, Alibaba’s Qwen, or Mistral’s Ministral difficult. The only usage feedback mentioned comes from Qwant, which already uses Olala mini for its quick responses and claims to get answers 20% faster with costs divided by three. This data should be taken with caution: Qwant was acquired by Synfonium, a group founded by Octave Klaba, which limits the neutrality of the example.

The model is based on an architecture developed by Dragon LLM, formerly Lingua Custodia, a Parisian startup acquired by OVHcloud in March and renamed OVHai LLM. Fundamentally, the group does not claim to catch up with the best models from Anthropic or OpenAI. It rather aims to be a European player capable of offering its own open models while continuing to run others’, including Chinese ones. The strategy is more pragmatic than spectacular. And in AI, that’s often where things get interesting.

Key points

  • Olala mini was presented on November 17, 2026.
  • OVHcloud announces 7 billion parameters, with 1 billion activated.
  • The model runs starting at 8 GB of graphics memory.
  • Qwant claims 20% faster responses and costs divided by three.
  • OVHcloud claims 1.7 million customers and 46 datacenters.

By the numbers

  • 7 billion parameters — total capacity of Olala mini, announced by OVHcloud in November 2026.
  • 1 billion active parameters — portion activated per request, according to OVHcloud.
  • 32,000 tokens — context window announced for the first version.
  • 46 datacenters — OVHcloud’s infrastructure cited in the announcement.
  • 1.7 million customers — commercial base highlighted by OVHcloud.

To read