Eigen RadarAI
Analysis

OVHcloud schedules Olala mini weights for 19 November

French cloud company OVHcloud has announced its first language-model family, beginning with Olala mini. Model weights and supporting documents are due on 19 November under an Apache 2.0 licence. Public European supercomputers supplied training capacity, while the search engine Qwant is already using the model for short answers. Later versions are planned for longer inputs and enterprise tasks.

Artificial Intelligence··Evening
An open supercomputer compute blade with copper heat-spreaders and coiled fiber-optic cables on a brightly lit workbench, illustrating the European supercomputing hardware used to train OVHcloud's Olala mini model.

A small model with selectively active components

OVHcloud, a cloud services company, announced Olala mini 7A1B on 17 November as the first member of its language-model family. The model has 7 billion parameters, the numerical values learned during training, with 1 billion active for a request. Its planned Apache 2.0 distribution includes downloadable weights. The official launch is scheduled for 19 November.[1], [2]

Each layer contains 256 specialist components, with only part of the model activated for an individual request. The initial context window, which sets the amount of text handled at once, is thirty-two thousand tokens, or text units. OVHcloud combines Differential-TPA and Mamba-3 in the model’s hybrid attention architecture.[1]

Public supercomputers supplied training capacity

Training used Jupiter, MeluXina and Jean Zay, three European public supercomputers. OVHcloud values the computing hours supplied through public support at €7 million. The search engine Qwant has already adopted Olala mini for its short-answer feature. The company plans to publish the pretrained base model and technical report alongside the weights, through the Hugging Face model-sharing platform, together with documentation addressing the AI Act, the European law on artificial intelligence.[1]

Later versions target longer inputs and workplace tasks

A version 1.1 planned before the end of the year would expand the context window to one hundred twenty-eight thousand tokens and improve mathematical reasoning. OVHcloud’s later 2.0 plans cover enterprise agents, cloud operations and cybersecurity. The company also intends to use locally running models in an encrypted version of its workspace software, allowing processing directly on the workstation. These are announced development plans for the family.[1]

References

  1. News sourceOVHcloudOVHcloud announces Olala mini ahead of 19 November launch↩1↩2↩3↩4
  2. News sourceActuIAOVHcloud unveils Olala mini, its first LLM: speeds, no scores↩