Olala mini 7A1B: OVH launches its first LLM, trained in Europe and open to all
OVH unveils Olala, its first family of language models. The first in the line, Olala mini 7A1B, is a compact model built from scratch by OVHai LLM, OVH’s agentic lab. It will be presented in detail on November 19 at the OVHcloud AI Summit, at the Carrousel du Louvre in Paris.
A compact model that runs anywhere
Olala mini has 7 billion parameters, only one billion of which are active, spread across 256 experts per layer, with a brand-new hybrid attention combining Differential-TPA and Mamba-3. It runs on a GPU with at least 8 GB of memory, reaches 160 tokens per second on an RTX A4500, and 50 on a standard CPU, with no GPU at all.
- Parameters
- 7 billion, 1 billion active
- Context
- 32,000 tokens
- Format
- BF16
- License
- Apache 2.0, on Hugging Face
Trained in Europe, already in production
Olala mini was trained on the European public supercomputers Jupiter, MeluXina and Jean Zay, the equivalent of €7 million in compute hours. This support made it possible to build the OVHai LLM team. The model is released with its weights, its pre-trained base model, its technical report and its EU AI Act compliance documentation. Qwant has already adopted it for its flash answer feature: responses are 20% faster and costs are divided by three.
What’s next: v1.1, then v2.0
Version 1.1, expected by the end of the year, will extend the context window to 128,000 tokens and strengthen mathematical reasoning. Version 2.0 will exceed 100 billion parameters and will come as models specialized in three areas: enterprise agentic AI, cloud DevOps and cybersecurity.
“We cannot simply rely on the generosity of certain labs to share their LLMs as open source in the years to come.”