OVHcloud's Olala: a small open French model, measured by OVH against Luciole, Liquid AI, Apertus, and EuroLLM

OVHcloud's Olala: a small open French model, measured by OVH against Luciole, Liquid AI, Apertus, and EuroLLM

On November 19, 2026, OVHcloud released the weights of Olala mini 7A1B (7 billion parameters, 1 billion active, Apache 2.0 license) and a technical report comparing it to 18 open models, including Luciole, Apertus, and EuroLLM. The model ranks seventh according to OVH, with a very interesting report that lists its own flaws.

In short

  • Olala mini 7A1B is a mixture of experts with 7 billion parameters, of which 1 billion are active per token, trained on 12.8 trillion tokens on three European supercomputers. The weights are released under the Apache 2.0 license.
  • In OVH's report, it ranks seventh out of 20 models (average of 68.0), behind six models, five of which are dense. As a base model, it outperforms Luciole-23B. It is strong in instruction following and weak in competition mathematics.
  • The measurements are OVH's. I have not found any independent evaluation.
  • Olala, Luciole, and Liquid AI models rely on three different architecture families.
ovhcloudsouveraineteopen-weightseurope

This is a short excerpt.

The full analysis and podcast are for subscribers. The account is free.

Sign in

No account yet?

← All the watch