OVHcloud releases Olala mini, an open model made only in Europe
The model has 7 billion parameters and operates on a standard CPU at 50 tokens for each second.
OVHcloud, a European cloud company, launched Olala, a family of open source AI models made only on European computers. The first model, Olala mini 7A1B, has 7 billion parameters, which are the values that a model learns in training. Chief executive Octave Klaba tells that OVHcloud cannot rely on some labs to share open source models in the years to come.
What is new
The model has a new structure for fast work with less memory. It is free with the Apache 2.0 licence on Hugging Face. The release has the model weights, technical specifications and documents for the EU AI Act.
The numbers
OVHcloud tells that the model makes 50 tokens, which are small pieces of text, for each second on a standard CPU. The French search engine Qwant uses it. Qwant reportedly gets answers 20% faster at one third of the cost.
What is not known
The context window, the text that the model reads at one time, is 32,000 tokens. The source gives no quality test results and no other models to compare.
Insight: Local open models reduce the need for big labs
Klaba tells that the cost of training fell by 20 times in three years. In November 2026 a researcher in India also released small Odia language models, with no test data. This reads as a race to control models near the user, since regional builders do not wait for large labs to share.
Sources
Posted