What OVHcloud announced
OVHcloud announced Olala mini 7A1B on 17 November 2026, the first language model in the Olala family, developed and trained from scratch by the French company. According to the source, the official presentation is scheduled for 19 November at the OVHcloud AI Summit in Paris, and distribution will take place in open source via Hugging Face under the Apache 2.0 license. The move is significant because it marks the cloud provider's entry into direct model design, not only the provision of infrastructure and compute capacity to run them.
The model was created with the stated goal of reducing the hardware resources required to use an LLM, in order to facilitate its adoption on local systems. The announced strategy includes agentic AI applications, developer tools and cybersecurity solutions.
Architecture and stated figures
Olala mini uses a mixture-of-experts architecture with 7 billion total parameters, of which about 1 billion are activated during processing, with 256 experts per layer. This approach makes it possible to select only a subset of components for each processing run, reducing operations compared with a model that uses all parameters. The system also incorporates a hybrid attention mechanism combining Differential-TPA and Mamba-3. The initial context window is 32 thousand tokens.
On the performance front, OVHcloud claims that the model can run on a GPU with 8 GB of memory, at 160 tokens per second on an Nvidia RTX A4500 and 50 tokens per second using only a CPU. However, the source notes that these figures require clarification: in BF16 format, the 7 billion parameters alone would occupy about 14 GB, and it has not been clarified which quantization or memory management techniques allow execution on an 8 GB GPU. There is also a lack of sufficient public benchmarks to compare answer quality with similarly sized models.
Training, license and first production use
Development was supported by access to three European public supercomputing infrastructures, Jupiter, MeluXina and Jean Zay, equipped with Nvidia GH200, A100 and H100 GPUs. According to OVHcloud, the resources made available are equivalent to about €7 million in compute hours: this is the value attributed to use of the infrastructures, not necessarily the total cost of the model. The project also led to the creation of OVHai LLM, a group lab dedicated to language models and agentic AI. Distribution will include weights, a pre-trained base version, technical report and documentation on AI Act obligations.
The first operational use announced concerns Qwant, which has integrated Olala mini into AI-generated quick answers. OVHcloud claims a 20% improvement in generation times and a reduction of operating costs to one third compared with the previous solution. These are results communicated by the company, not yet accompanied by an independent evaluation. The next milestones have also been outlined: version 1.1 by the end of 2026 should bring the context to 128 thousand tokens, while version 2.0, with no date, should exceed 100 billion parameters with specialized models for agentic AI, cloud DevOps and cybersecurity.
