OVHcloud announced Olala, its first family of AI models. The first release, Olala mini 7A1B, uses a 7 billion parameter architecture. The model will be available as open source on Hugging Face under the Apache 2.0 license.
Training and launch
Olala mini was trained on three European public supercomputers: Jupiter, MeluXina, and Jean Zay. The public support represented €7 million in computing hours. The model's official launch will take place at the OVHcloud AI Summit on November 19.
Qwant adoption
Qwant already uses Olala mini for its flash response feature. Responses became 20% faster, according to OVHcloud. Operating costs fell by two-thirds compared with the previous model.
Model design
Olala mini has one billion active parameters across 256 experts per layer. It uses hybrid attention with Differential-TPA and Mamba-3. Its current context window holds 32,000 tokens.
Version 1.1 is expected by the end of the year. It will raise the context window to 128,000 tokens and improve mathematical reasoning. Version 2.0 plans specialized models for enterprise agentics, cloud DevOps, and cybersecurity, and it is expected to exceed 100 billion parameters.