AI AI Sovereignty

OVHcloud launches Olala mini 7A1B, its first open AI model

French cloud giant debuts open-weights LLM family with hybrid attention architecture and already has Qwant as its first production user

• Updated 23 hours ago
OVHcloud revela o Olala mini 7A1B, primeiro modelo da família de LLMs abertos que será apresentada oficialmente no AI Summit, em Paris
OVHcloud revela o Olala mini 7A1B, primeiro modelo da família de LLMs abertos que será apresentada oficialmente no AI Summit, em Paris (Ilustração editorial / NEWS by Box4Dev)

Spec Sheet

  • Olala mini 7A1B is the first model in the family, featuring a hybrid attention architecture inherited from Dragon LLM
  • The official launch takes place on November 19, at the OVHcloud AI Summit, at the Carrousel du Louvre in Paris
  • Qwant reports 20% faster responses and costs three times lower with the OVHcloud model

OVHcloud unveiled on Tuesday (17) Olala, its first family of large language models. It is the most concrete step yet by the French cloud giant to stop being merely the infrastructure provider where third-party AI runs. The official launch is set for November 19, during the OVHcloud AI Summit, at the Carrousel du Louvre in Paris.

The first member of the family is Olala mini 7A1B, a compact model aimed at relatively light workloads. OVHcloud did not disclose a full spec sheet: the company limits itself to describing the model as the first of a lineage built on a novel architecture dubbed hybrid attention.

A small model with Parisian heritage

The technical foundation of Olala mini was not born inside OVHcloud. It comes from Dragon LLM, formerly Lingua Custodia, a Paris-based startup acquired by the group in March and renamed OVHai LLM. The 7A1B nomenclature follows the convention for sparse-architecture models, which separates total parameters from active parameters per inference — a design meant to lower execution costs.

"Hybrid attention" combines different attention mechanisms within the same network, reducing the volume of memory and processing required in long contexts. It is exactly the kind of optimization that interests a cloud provider: every percentage point of efficiency becomes margin at datacenter scale.

The stated ambition is modest, and the company does not hide it. OVHcloud does not intend to beat the frontier models from Anthropic or OpenAI — it wants to position itself as a European player capable of offering its own open-weights models, while continuing to host other companies' models, including Chinese ones.

The trump card the labs don't have

The timeline was moved up. In June, founder and CEO Octave Klaba had already announced the intention to train its own models, but spoke of a launch in early 2027. Tuesday's reveal, on November 17, 2026, shortened the deadline by about a quarter.

OVHcloud's differentiator is not in the lab, but in distribution: 1.7 million customers and 46 datacenters. No frontier AI lab operates an installed base of that kind, which makes it possible to bundle model, inference and hosting into a single contract — a direct sales argument for European companies under pressure from compliance and data sovereignty.

Until now, Mistral AI and its frontier models concentrated practically all French expectations in the field. With Olala, France stops depending on a single domestic model supplier.

Qwant tests it, but the test has an owner

The only public sign of performance comes from Qwant, which already uses Olala mini in its flash answers. The French search engine reports 20% faster responses and costs cut by three. The number, however, must be read with caution: Qwant was bought by Synfonium, a group founded by Octave Klaba himself.

In other words, the first reference customer is also a company tied to OVHcloud's controller. This is not an independent test, and the company itself has not published benchmarks that would allow comparison.

The gap that still weighs

Without official numbers, it is impossible to place Olala mini 7A1B against Gemma from Google, Qwen from Alibaba, or Ministral from Mistral — its direct competitors in that tier of compact models. In open weights, that gap is especially sensitive: developers choose a model based on public evaluation, not on a press release.

For the European market, the absence of benchmarks is the main brake on adoption. OVHcloud's privileged distribution solves the commercial problem, but it does not replace the technical evidence that the open source ecosystem uses as its yardstick.

Practical Impact

On the technical side, a compact model with hybrid attention attacks precisely the cost line that weighs most in inference: context memory and response time. If the reported efficiency holds up under real workloads, the model serves as a cheap layer for search, classification and customer service, leaving frontier models for complex tasks.

On the operational side, OVHcloud shifts the AI dispute from the research field to the infrastructure field: selling its own inference to a base of 1.7 million customers and 46 datacenters. Success depends less on surpassing Mistral in quality and more on convincing the market that open weights, without public benchmarks, are reliable enough for production.

Primary Source