A compact model called Olala marks OVHcloud’s entry into LLMs

The Olala family starts with a 7-billion-parameter model. Already used by Qwant, it is set to be released under the Apache 2.0 license.

The Qwant search engine’s flash answers already use the first language model in the Olala family, developed by OVHcloud. According to the provider, the deployment delivers answers 20% faster and cuts operating costs to one-third of their previous level. The model it replaced is not named, and the press release does not detail the conditions of the comparison.

This initial deployment accompanies the announcement of Olala mini 7A1B, whose official launch is scheduled for November 19 at the OVHcloud AI Summit. The company plans to release it on Hugging Face under the Apache 2.0 license, alongside its weights, pretrained base model, technical report, and AI Act compliance documentation. Seven billion parameters, one billion active

The 7A1B designation captures the model’s central distinction: seven billion parameters in total, with one billion activated to process a request. Its architecture includes 256 experts per layer and combines Differential-TPA with Mamba-3. This selective approach reduces the computation required for each generation without reducing the model’s total size to one billion parameters.

For inference, the company reports a throughput of 160 tokens per second on an Nvidia RTX A4500 graphics card and 50 tokens per second on a CPU, without specifying which processor. It also says the model can run on a GPU with 8 GB of memory. The numerical precision and loading configuration used to reach that threshold are not detailed, although BF16 is listed as the model’s format. Its context window reaches 32,000 tokens, the amount of text the system can take into account within a single interaction. Seven million euros in publicly funded compute time

Training used three European public supercomputers: Jupiter, MeluXina, and Jean Zay. The hardware cited includes GH200, A100, and H100 GPUs. According to OVHcloud, access to these resources represents the equivalent of €7 million in compute time. That figure values the support received; it is not a total training budget disclosed by the company.

The work led to the creation of the OVHai LLM lab. For OVHcloud CEO Octave Klaba, developing this expertise addresses both internal needs and customer demand: having in-house models without depending on other labs’ decisions about whether to release theirs openly. From desktop computers to cybersecurity

Version 1.1 is expected before the end of 2026, with a context window expanded to 128,000 tokens and planned improvements in mathematical reasoning. These are upcoming changes, separate from the capabilities presented for the initial version.

Longer-term development is intended to support enterprise agents and OVHai Workspace, whose encrypted version will require models that run directly on users’ computers. Two other areas are planned: tools for cloud development and operations teams, and defensive and offensive cybersecurity. For version 2.0, the company plans to exceed 100 billion parameters, without providing a release date.