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

Olala mini is based on an architecture of 7 billion parameters (7B) with 1 billion activated, distributed among 256 experts per layer, and hybrid attention with Differential-TPA and Mamba-3, in BF16 format. Only a fraction of the model is used in each query, making it faster and consuming less memory. As a direct consequence, the model runs on a GPU with only 8 GB of memory, reaches 160 tokens per second on an RTX A4500 and, even without a GPU, on a standard processor, reaches 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 license.

Already up and running on Qwant

Olala mini has been trained on three European public supercomputers – Jupiter, MeluXina and Jean Zay – equipped with GH200, A100 and H100 GPUs. This support, which represents the equivalent of 7 million euros in calculation hours, has been decisive, and OVH would like to express its gratitude to the authorities. The public initiative has proven to be a success in more ways than one: it has given rise to an exceptional team in artificial intelligence and has given rise to the OVH agent laboratory: OVHai LLM. Olala mini is launched with its weights, its pre-trained base model, its technical report and its documentation in compliance with the AI ​​Law. Qwant has already adopted it for its instant responses feature, with significant benefits: 20% faster responses and running costs reduced by a third compared to the previously used model.

Olala will promote OVHai Workspace

Expected by the end of the year, version 1.1 will expand the context window to 128,000 tokens and improve its performance in mathematical reasoning. With version 2.0, OVHai LLM aims to develop specialized models in three areas. First, agentic AI for enterprises: over time, Olala will power OVHai Workspace, the encrypted version of which will require models capable of running locally, directly in the workplace. Secondly, DevOps in the cloud, with versions adapted to developers, architects and operations teams that guarantee service continuity 24 hours a day. Finally, cybersecurity, which covers both defensive and offensive aspects. Olala v2.0 will exceed 100 billion parameters.

“AI is strategic for OVH. In addition to offering AI services to our clients, we must be able to create models both for our own internal needs and for those of our clients. We cannot simply rely on the generosity of certain laboratories to share their LLM models as open source in the coming years. The know-how of model training has become widely widespread and the cost of training has been reduced by one-twentieth in three years. It is now possible to create LLM models and, at the same time, make this activity be profitable,” said Octave Kalba, CEO of OVHcloud.