Summary
- Olala mini has seven billion total parameters, with approximately one billion active during inference.
- European public supercomputers supported training, while an Apache 2.0 release is planned for 19 November.
- Qwant reports faster answers and lower operating costs, although independent performance evidence remains limited.
French cloud provider OVHcloud has developed its first large language model, moving beyond the provision of artificial intelligence infrastructure into the design and training of its own AI systems.
The company introduced Olala mini 7A1B on 17 November, ahead of a formal presentation at its AI Summit on 19 November. The model is designed to reduce the computing resources required for inference, with an architecture that activates only part of its total parameter collection during processing.
Olala mini contains approximately seven billion parameters, of which around one billion are active during an individual operation. OVHcloud says the model can run on a graphics processor with as little as 8GB of memory and will be released through Hugging Face under the Apache 2.0 licence.
French search engine Qwant is already using the technology for its flash answer service. According to OVHcloud, the deployment has improved response times by 20% and reduced operating costs to one-third of those associated with the previous model.
Selective computation reduces processing demands
Language models rely on large collections of numerical parameters developed during training. Increasing their size can improve their ability to perform certain tasks, but usually creates additional demands for computing power, memory and infrastructure.
Olala mini uses a mixture-of-experts architecture, distributing its parameters between specialist components rather than activating every parameter for each operation.
Its approximately one billion active parameters are drawn from a larger collection of seven billion, with OVHcloud reporting 256 experts per layer.
The architecture also incorporates Differential-TPA attention and Mamba-3 components, combining different methods for processing information and maintaining context.
OVHcloud says the model achieves 160 tokens per second on an NVIDIA RTX A4500 and 50 tokens per second on a conventional processor without a dedicated GPU.
These are company-reported results, and the precise hardware configurations, workloads and measurement conditions require independent examination before broader performance comparisons can be made.
The claimed 8GB deployment requirement also needs technical clarification. Seven billion parameters stored directly in BF16 format would occupy approximately 14GB before accounting for other memory requirements, although mixture-of-experts models can use specialised memory management and storage techniques.
OVHcloud has not yet published sufficient technical detail to establish how its stated minimum configuration is achieved or which operating conditions apply.
The model currently supports a context window of 32,000 tokens. That determines how much information can be supplied within an interaction, although effective use of longer documents depends on the model’s ability to retrieve and interpret the relevant material.
Lower memory requirements could make deployment practical on a wider range of equipment, reducing dependence on expensive GPU infrastructure for particular applications.
However, efficiency does not establish suitability for every task. Organisations must evaluate whether the model provides acceptable accuracy, reasoning capability and reliability for their intended workloads.
European supercomputers support development
Olala mini was trained using three European public supercomputers: Jupiter in Germany, MeluXina in Luxembourg and Jean Zay in France.
The training infrastructure included NVIDIA GH200, A100 and H100 processors, with OVHcloud estimating that the public computing contribution was equivalent to approximately €7 million in machine time.
That amount represents the estimated value of computing resources rather than a direct cash investment in the company.
Publicly supported high-performance computing facilities can give European developers access to equipment that would otherwise require substantial private investment.
Model training nevertheless involves more than acquiring computing capacity. Developers also need suitable datasets, engineering expertise, evaluation methods and systems capable of supporting the resulting software in production.
OVHcloud plans to release Olala mini’s model weights, pretrained base model, technical report and documentation relating to the EU AI Act.
The proposed Apache 2.0 licence would permit broad commercial use and modification, subject to its terms and any other applicable legal obligations.
Publication will allow independent developers to inspect the model, reproduce performance tests and assess whether its capabilities justify replacing existing alternatives.
Qwant provides an early commercial application
Qwant’s adoption provides an initial example of how OVHcloud expects the model to be used commercially.
The search engine uses Olala mini to generate concise answers to user queries, an application in which response speed and the cost of processing large volumes of requests are commercially important.
OVHcloud reports that the system produces answers 20% faster while reducing running costs by approximately two-thirds compared with the previous model.
The comparison does not identify the replaced model or provide a detailed breakdown of the implementation, limiting the conclusions that can be drawn about the contribution of Olala mini itself.
Qwant is also not an entirely independent reference customer. The search engine is controlled through Synfonium, a business associated with OVHcloud founder Octave Klaba and his family’s investment interests.
Its use of Olala mini demonstrates deployment within a working service, but independent customer adoption and independently reproducible performance results will provide stronger evidence of commercial competitiveness.
Other enterprise applications may create different requirements. Document analysis, customer support, software development and specialist business workflows can demand greater reasoning capability or longer context handling than a search answer service.
Performance will therefore need to be assessed against the requirements of each application rather than inferred from the Qwant results.
A broader model roadmap takes shape
OVHcloud is already planning additional versions of Olala.
Olala 1.1 is expected before the end of 2026, with a proposed context window of 128,000 tokens and improvements to mathematical reasoning.
A larger version 2.0 is intended to exceed 100 billion parameters, supporting more specialised applications involving enterprise agents, cloud operations and cybersecurity.
These capabilities remain on the company’s development roadmap and should not be confused with the initial model announced in November.
OVHcloud is also considering applications within OVHai Workspace, its AI collaboration environment. Future deployments could include configurations designed to operate locally on customer equipment.
Local inference may provide greater direct control over certain information flows, although organisations must still secure their devices, manage model updates and establish appropriate controls over the information supplied to AI applications.
By developing its own models, OVHcloud gains an additional option for building services without depending exclusively on external AI suppliers and their licensing or commercial terms.
That independence introduces new responsibilities for model maintenance, security, evaluation and technical support.
The company’s ability to compete will depend on whether the resulting models provide sufficient capability at attractive operating costs, rather than on European infrastructure provenance alone.
The planned 19 November release will make the technical evidence more accessible. Model weights, benchmark configurations and independent testing should clarify whether Olala mini’s reported efficiency translates into a practical advantage across different enterprise workloads.












