OVHcloud announces Olala mini, a compact model aimed at local AI

OVHcloud announced Olala mini 7A1B on November 17, introducing a compact language model with seven billion total parameters and one billion active parameters. The company schedules its formal launch for the OVHcloud AI Summit on November 19.
The announcement describes a hybrid architecture combining Differential-TPA attention and Mamba-3, with 256 experts per layer and a 32,000-token context window. OVHcloud says the model can run with 8 GB of GPU memory and reports 160 tokens per second on an RTX A4500. Those are company figures, not independently reproduced results.
OVHcloud says Qwant already uses the model for flash answers. Training used the European public supercomputers Jupiter, MeluXina and Jean Zay, with computing support valued at €7 million.
The company specifies Apache 2.0 licensing and a Hugging Face distribution plan. Its announcement uses both current and future wording for availability; a downloadable model repository was not independently verified for this report.
A 128,000-token context window is planned for version 1.1 by year-end, not a current feature. The announcement matters for local AI deployments seeking compact models, while availability and performance still need checking against the eventual model files and documentation.
Sources and citations
Revision history
Published .