OVHcloud has released Olala mini 7A1B, the first model in a new family of large language models, and says it was trained entirely on three European public supercomputers: Jupiter, MeluXina and Jean Zay. Those machines use GH200, A100 and H100 GPUs. The model will be published on Hugging Face under the Apache 2.0 licence.
Olala mini has 7 billion parameters, of which 1 billion are active, spread across 256 experts per layer, with hybrid attention that combines Differential-TPA and Mamba-3 in BF16 format. Because only a fraction of the model is used for each request, OVHcloud says it is faster and needs less memory. It runs on a GPU with as little as 8GB of memory, reaching 160 tokens per second on an RTX A4500, and on a standard processor without a GPU at 50 tokens per second. The context window is currently 32,000 tokens.
OVHcloud is releasing the weights, the pre-trained base model, a technical report and its AI Act compliance documentation. French search engine Qwant already uses the model for its flash answer feature and reports answers 20% faster and running costs divided by three compared with the model it used before.
Octave Klaba, chief executive of OVHcloud, gave the strategic reasoning.
A version 1.1 of Olala mini is expected by the end of the year, extending the context window to 128,000 tokens and improving mathematical reasoning. Version 2.0 will aim at three areas: agentic AI for businesses, which would eventually power OVHai Workspace; cloud DevOps for developers, architects and operations teams; and cybersecurity, covering both defensive and offensive uses. OVHcloud says version 2.0 will exceed 100 billion parameters.
To stay across the latest in cloud, AI and enterprise tech analysis from Compare the Cloud, subscribe to our weekly newsletter at https://www.comparethecloud.net/newsletter