OVHcloud Introduces Olala Mini 7A1B as Its First Open-Source LLM

Unite.AI
Read full story· 4 min
OVHcloud launched Olala mini 7A1B, its first open-source large language model that runs on GPUs with 8 GB memory. Trained on European supercomputers, it features hybrid attention for efficiency and is already deployed by Qwant.

What happened

On November 17, 2026, OVHcloud introduced Olala mini 7A1B, its first open-source large language model. The model is designed to operate on GPUs with 8 GB of memory and will be released on Hugging Face under the Apache 2.0 license. Olala mini 7A1B was trained on European supercomputers and is already being used by Qwant for its flash-answer feature. OVHcloud plans to present the model at its AI Summit on November 19, 2026.

Key details

  • Olala mini 7A1B has a 7-billion-parameter architecture, with only 1 billion parameters active per request.
  • The model uses hybrid attention combining Differential-TPA and Mamba-3, and processes up to 32,000 tokens per sequence.
  • Qwant reports 20% faster answers and one-third the running cost compared to its previous model.
  • Training used three European supercomputers: Jupiter, MeluXina, and Jean Zay, with €7 million worth of computing hours.
  • Future versions are planned, including a 1.1 update by the end of 2026 and a 2.0 version exceeding 100 billion parameters.

Why it matters

European companies and public sector organizations now have access to an open-source language model optimized for lower hardware requirements, reducing costs and broadening deployment options. Qwant, as an early adopter, benefits from faster response times and significantly lower operational expenses. OVHcloud customers and partners will soon face new choices as the company expands the Olala model family into business, DevOps, and security applications, with compliance documentation included for EU AI Act requirements.


More in LLM & Text Generation