N環球 N GLOBE
France Technology

France's OVHcloud unveils Olala mini, its first open-source LLM: 7B parameters, 1B active per token

· N GLOBE

French cloud provider OVHcloud on November 17 announced Olala, its first family of large language models. The first model, Olala mini 7A1B, activates only about 1 billion of its 7 billion parameters per token and can run on an 8GB GPU.

French cloud provider OVHcloud announced on November 17 in Roubaix the launch of Olala, its first family of large language models built in-house. The first model, Olala mini 7A1B, uses a 7-billion-parameter architecture but activates only about 1 billion parameters per request, and the company says it can run on a graphics card with as little as 8GB of memory, or even on a standard processor. The model will be released as open source on Hugging Face under the Apache 2.0 licence, with the official launch set for the OVHcloud AI Summit on November 19.

Mixture-of-experts design, 256 experts per layer

Olala mini uses a mixture-of-experts (MoE) architecture with 256 experts per layer, activating only a subset at a time so active parameters stay near 1 billion. Its attention mechanism combines Differential-TPA and Mamba-3, computed in BF16 format. The context window is 32,000 tokens; OVHcloud says version 1.1, due by the end of the year, will extend it to 128,000 tokens and improve mathematical reasoning. The company reports speeds of 160 tokens per second on an RTX A4500 GPU and 50 tokens per second on a CPU, both vendor-measured figures.

Trained on three European supercomputers; Qwant already uses it

The model was trained from scratch by the OVHcloud lab on three European public supercomputers — Jupiter in Germany, MeluXina in Luxembourg and Jean Zay in France — equipped with GH200, A100 and H100 GPUs. French search engine Qwant already uses Olala mini for its flash-answer feature; OVHcloud says it is 20 percent faster and costs one-third as much to run as the previous model. Founder Octave Klaba said developers will be able to download the model and run it on their own machines.

No benchmark scores yet; open-source details due November 19

Notably, OVHcloud has not published any public benchmark scores for Olala mini, so there is no way yet to judge it against open models of a similar size — a gap flagged by French tech publication Numerama. The press release also gives no Hugging Face repository address, and as of November 17 the model had not appeared on OVHcloud's OVHaiLLM profile. The release package will include the model weights, the pre-trained base model, a technical report and EU AI Act compliance documentation, with details to come at the November 19 launch. (Sources: OVHcloud press release; AFP; Unite.AI; Numerama)

(Illustration: AI data centre. Image: 618Media.)

💬 Comments

No comments yet. Be the first!