AI Aulis

AI NEWS IN ENGLISH · 19 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

38 stories · next issue tomorrow morning · How the paper is made →

← Issue 19 November 2026Paper

OVHcloud unveils its first language model, the 7-billion-parameter Olala mini

AI-generated contentWritten by AI

m.ithome.com, corporate.ovhcloud.com, next.ink · 5 sources · Sources published: 17 November 2026

French cloud company OVHcloud will release Olala mini 7A1B on Hugging Face under the Apache 2.0 licence. The company says the model runs on an 8 GB GPU, but Next.ink calculates that the parameters alone take up 14 GB in BF16 format.

Olala mini is the first model in OVHcloud’s new Olala family, the company’s first series of large language models. OVHcloud’s press release says the model will be released as an open-weight model. Alongside the weights, the company will release the pre-trained base model, a technical report and compliance documentation for the EU AI Act. The official launch is at the OVHcloud AI Summit on 19 November.

Olala mini is a mixture-of-experts (MoE) model, meaning it uses only part of its network at a time. One billion of its 7 billion parameters are active. In each layer, the active parameters are spread across 256 experts. OVHcloud says this architecture makes the model faster and reduces its memory requirements. French publication Next.ink points out that the architecture speeds up computation but does not reduce memory requirements, because all 7 billion parameters still have to be loaded into memory.

OVHcloud says the model runs on a GPU with at least 8 GB of memory. It generates 160 tokens per second on an RTX A4500 card and 50 tokens per second on a processor alone, the company says. Next.ink calculates that the parameters in BF16 format take up 14 GB on their own, and the context needs additional space. The model fits into 8 GB when quantised, with weights stored at 8- or 4-bit precision, which brings them to 7 or 3.5 GB, the publication said. Next.ink also notes that the RTX A4500 used for the speed test has 20 GB of memory, and that OVHcloud does not say which processor was used to measure the 50 tokens per second.

Apart from speed, OVHcloud has not published any other benchmark results, so the model’s quality cannot yet be assessed, Next.ink reported. The first use case comes from OVHcloud itself. Its press release says the search engine Qwant uses the model for its flash answers. Answers arrive 20 per cent faster and running costs are a third of those of the model it used before. Next.ink reports that Qwant has so far used Mistral’s models and that the search engine belongs to a company owned jointly by the Klaba family and the Caisse des dépôts. OVHcloud’s CEO is Octave Klaba.

The model was trained on three European public supercomputers: Jupiter, MeluXina and Jean Zay. The press release says public support was equivalent to €7 million in computing time. China’s ithome.com (IT之家) reported the sum as the total cost of training. OVHcloud has not disclosed its own investment, Next.ink said.

The context window is currently 32,000 tokens. Version 1.1, expected by the end of the year, will extend it to 128,000 tokens and improve mathematical reasoning. No timetable has been given for version 2.0. The plan is to split it into models specialised in enterprise agent use, DevOps work for cloud services and cybersecurity, and the company says it intends to increase the parameter count to more than 100 billion.

Report an error in this story