With Olala mini 7A1B, OVHcloud releases the first model in its new Olala family of language models. The model is open source and already runs at search engine Qwant.
OVHcloud has announced Olala, a family of large language models that the company trains itself in Europe. Olala mini 7A1B is the first model. The official launch follows on 19 November at the OVHcloud AI Summit.
The model uses seven billion parameters, of which one billion are active, spread across 256 experts per layer. For each query, it activates only part of the model. That makes it faster and less memory-hungry.
Runs on modest hardware
Olala mini works on a GPU with 8 GB of memory and reaches 160 tokens per second on an Nvidia RTX A4500. Without a GPU, the model reaches 50 tokens per second on a regular processor. For now, the context window is 32,000 tokens.
The attention combines Differential-TPA and Mamba-3, in BF16 format. OVHcloud publishes the model on Hugging Face under the Apache 2.0 licence, together with the weights, the pretrained base model, a technical report and the documentation for the AI Act.
Qwant already uses Olala mini for its flash answer feature. According to OVHcloud, the search engine now responds twenty percent faster and operating costs have dropped to a third of those of the previous model.
Trained on European supercomputers
OVHcloud trained the model on the public supercomputers Jupiter, MeluXina and Jean Zay, using GH200, A100 and H100 GPUs. That support is equivalent to 7 million euros in compute hours. The initiative also gave rise to OVHcloud’s Agentic Lab, OVHai LLM.
Octave Klaba, CEO of OVHcloud, explains the choice: “We cannot keep relying on the generosity of certain labs to share their language models as open source in the coming years.” According to him, the knowledge to train models is widely available, and training costs have fallen by a factor of twenty in three years.
Version 1.1 and 2.0
Version 1.1 is due by the end of the year, with a context window of 128,000 tokens and better mathematical reasoning. Version 2.0 will have more than a hundred billion parameters. OVHai LLM then wants to build specialised models for AI agents in enterprises, cloud DevOps and cybersecurity.
In the longer term, Olala is also meant to power OVHai Workspace. The encrypted version of that requires models that run locally on the workstation. The company has not yet disclosed what the model will cost when used via the OVHcloud platform.
