OVH announces Olala mini 7A1B

OVH announces Olala, its first family of LLMs. The first member of this family, Olala mini 7A1B, is a compact model built on a novel architecture. It demonstrates that European companies can innovate and train agentic LLMs from scratch. Its official launch will take place at the OVHcloud AI Summit on 19 November, where OVH will present in detail all of its cloud and AI innovations for its customers.

Olala: an innovative model built on hybrid attention

Olala mini is based on a 7-billion-parameter architecture (7B) with one billion active parameters spread across 256 experts per layer, and hybrid attention combining Differential-TPA and Mamba-3 in BF16 format. Only a fraction of the model is used for each request, making it faster and less memory-hungry. As a direct result, the model runs on a GPU with as little as 8 GB of memory, reaching 160 tokens per second on an RTX A4500 and even without a GPU, on a standard processor at 50 tokens per second. Its context window currently reaches 32,000 tokens. Olala mini will be available as open source on Hugging Face under the Apache 2.0 licence.

Trained with European public support and already running at Qwant

Olala mini was trained on three European public supercomputers, Jupiter, MeluXina and Jean Zay, equipped with GH200, A100 and H100 GPUs. This support, equivalent to €7 million in computing hours, was decisive and OVH wishes to thank the public authorities. The public initiative has proved a success on more than one count: it has brought together an exceptional artificial intelligence team and given rise to OVH’s Agentic Lab, OVHai LLM. Olala mini is released with its weights, pre-trained base model, technical report and AI Act compliance documentation. Qwant has already adopted it for its flash answer feature with significant gains: answers 20% faster and running costs divided by three compared with the model previously used.

On the road to v1.1, then v2.0

Expected by the end of the year, version 1.1 will extend the context window to 128,000 tokens and improve its performance in mathematical reasoning. With version 2.0, OVHai LLM intends to develop models specialised in three areas. First, agentic AI for businesses: in time, Olala will power OVHai Workspace whose encrypted version will require models capable of running locally, directly on the workstation. Second, cloud DevOps with versions tailored to developers, architects and the operations teams who ensure round-the-clock service continuity. Finally, cybersecurity covering both defensive and offensive aspects. Olala v2.0 will exceed 100 billion parameters.

Octave Klaba, CEO of OVHcloud, said, “AI is strategic for OVH. As well as offering AI services to our customers, we must be able to create models for our own internal needs and for those of our customers. We cannot simply rely on the generosity of certain labs to share their LLMs as open source in the years to come. Know-how in model training has become widely accessible, and the cost of training has been divided by 20 in three years. It is now possible to create LLMs while making this activity profitable.”

Leave a Reply

Your email address will not be published. Required fields are marked *