OVHcloud announces first open-source model family, trained on European public supercomputers OVHcloud announced Olala mini 7A1B, its first open-source large language model, a 7-billion-parameter mixture-of-experts model with 256 experts per layer and 1 billion active parameters per request, trained from scratch on three EuroHPC supercomputers — Jupiter in Germany, MeluXina in Luxembourg and Jean Zay in France — using GH200, A100 and H100 GPUs, with OVHcloud valuing the public compute allocation at €7 million. Weights, a pre-trained base model, a technical report and AI Act compliance documentation will be published on Hugging Face under the Apache 2.0 licence, with an official launch at the OVHcloud AI Summit on 19 November at the Carrousel du Louvre in Paris. OVHcloud cites 160 tokens per second on an Nvidia RTX A4500 and 50 tokens per second on an unnamed standard CPU, but does not explain how the roughly 14 GB of BF16 weights fit on the 8 GB GPU it says the model can run on. OVHcloud announces first open-source model family, trained on European public supercomputers OVHcloud has announced Olala, its first family of large language models, starting with a compact 7-billion-parameter model it says was trained entirely on European public supercomputers. The announcement is the most concrete step yet by the European cloud provider toward its stated goal of AI sovereignty: building and controlling the full stack from datacentres to model weights. The first member, Olala mini 7A1B, uses a mixture-of-experts architecture with 256 experts per layer. Only one billion parameters are active for any given request, a design OVHcloud says reduces memory demand enough to run on a GPU with 8 GB of memory or on a standard CPU without a graphics card at all. The model combines Differential-TPA and Mamba-3 attention in BF16 format and offers a 32,000-token context window. Weights, a pre-trained base model, a technical report and AI Act compliance documentation will be published on Hugging Face under the Apache 2.0 licence. An official launch presentation is scheduled for the OVHcloud AI Summit on 19 November at the Carrousel du Louvre in Paris \ 1\ https://corporate.ovhcloud.com/en/newsroom/news/ovh-olala-mini . Training on public infrastructure OVHcloud states that Olala mini was trained from scratch on three EuroHPC supercomputers: Jupiter in Germany, MeluXina in Luxembourg and Jean Zay in France, equipped with GH200, A100 and H100 GPUs. The company values the public compute allocation at €7 million and thanks the public authorities for the support \ 1\ https://corporate.ovhcloud.com/en/newsroom/news/ovh-olala-mini . OVHcloud does not specify under what scheme the compute hours for Olala were allocated \ 4\ https://www.actuia.com/en/news/ovhcloud-unveils-olala-mini-its-first-llm-speeds-no-scores . Dragon LLM — a Paris-based startup OVHcloud acquired in March 2026 and renamed OVHai LLM — won the European Commission's Large AI Grand Challenge in June 2024, obtaining access to the Leonardo and Jupiter supercomputers \ 6\ https://www.hub-franceia.fr/wp-content/uploads/2025/11/FR-VF-Lancement-Dragon-15-Octobre-2025.pdf . The Grand Challenge award included 3.2 million compute hours on those machines, Silicon.fr reported \ 4\ https://www.actuia.com/en/news/ovhcloud-unveils-olala-mini-its-first-llm-speeds-no-scores . Dragon LLM had previously demonstrated a 3.6-billion-parameter model on Leonardo and Jupiter \ 4\ https://www.actuia.com/en/news/ovhcloud-unveils-olala-mini-its-first-llm-speeds-no-scores \ 6\ https://www.hub-franceia.fr/wp-content/uploads/2025/11/FR-VF-Lancement-Dragon-15-Octobre-2025.pdf . Olala mini's training used a different set of machines: Jupiter, MeluXina and Jean Zay, with only Jupiter in common \ 1\ https://corporate.ovhcloud.com/en/newsroom/news/ovh-olala-mini \ 4\ https://www.actuia.com/en/news/ovhcloud-unveils-olala-mini-its-first-llm-speeds-no-scores . The acquisition of Dragon LLM is the source of the hybrid attention architecture used in Olala \ 4\ https://www.actuia.com/en/news/ovhcloud-unveils-olala-mini-its-first-llm-speeds-no-scores . OVHcloud also acquired Seald end-to-end encryption in January 2026 and Gladia speech recognition in July 2026, folding all three into its AI Lab \ 5\ https://blog.ovhcloud.com/fr/posts/ai-summit-2026-sovereign-ai . Performance claims and gaps OVHcloud publishes two throughput figures: 160 tokens per second on an Nvidia RTX A4500 and 50 tokens per second on an unnamed standard CPU \ 1\ https://corporate.ovhcloud.com/en/newsroom/news/ovh-olala-mini . The RTX A4500 is a workstation card with 20 GB of memory, not the 8 GB minimum the company cites \ 4\ https://www.actuia.com/en/news/ovhcloud-unveils-olala-mini-its-first-llm-speeds-no-scores . In BF16, the full 7-billion-parameter model occupies roughly 14 GB of VRAM; OVHcloud does not explain how the weights fit on an 8 GB card, leaving open whether a quantised version or partial expert loading is intended \ 4\ https://www.actuia.com/en/news/ovhcloud-unveils-olala-mini-its-first-llm-speeds-no-scores . No benchmark scores — MMLU, GSM8K, HumanEval or any standard evaluation — are included in the announcement. The technical report, promised for release on 19 November, is expected to provide them. The only deployed example OVHcloud cites is Qwant, the French search engine, which has adopted Olala mini for its "flash answer" summaries. OVHcloud reports answers 20 % faster and running costs divided by three compared with Qwant's previous, unnamed model \ 1\ https://corporate.ovhcloud.com/en/newsroom/news/ovh-olala-mini . Qwant is not an independent customer: it is owned by Synfonium, which is 75 % held by Octave and Miroslaw Klaba OVHcloud's founder and his brother and 25 % by Caisse des Dépôts \ 4\ https://www.actuia.com/en/news/ovhcloud-unveils-olala-mini-its-first-llm-speeds-no-scores . The cost and speed gains describe this single integration, not a general comparison with other open models. Roadmap and strategy OVHcloud frames Olala as the first step in a broader programme. Version 1.1, expected before the end of 2026, will extend the context window to 128,000 tokens and improve mathematical reasoning. Version 2.0, undated, is planned to exceed 100 billion parameters with models specialised for business agentic AI, cloud DevOps and cybersecurity both defensive and offensive \ 1\ https://corporate.ovhcloud.com/en/newsroom/news/ovh-olala-mini \ 3\ https://www.unite.ai/ovhcloud-introduces-olala-mini-7a1b-as-its-first-open-source-llm . The models are intended to power OVHai Workspace, a collaborative platform with messaging, storage and videoconferencing that OVHcloud previewed in June 2026 \ 5\ https://blog.ovhcloud.com/fr/posts/ai-summit-2026-sovereign-ai ; an encrypted version will require models that run locally on workstations \ 4\ https://www.actuia.com/en/news/ovhcloud-unveils-olala-mini-its-first-llm-speeds-no-scores . CEO Octave Klaba linked the investment to strategic independence: "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" \ 1\ https://corporate.ovhcloud.com/en/newsroom/news/ovh-olala-mini . What is real and what is promised What is shipped today: a 7B/1B sparse model, trained on European public supercomputers via public compute support valued at €7 million, with Apache 2.0 weights and documentation promised for 19 November. What remains unverified: benchmark quality, actual memory footprint on an 8 GB GPU, CPU performance on specified hardware, and the training data summary required by the AI Act since August 2025 \ 4\ https://www.actuia.com/en/news/ovhcloud-unveils-olala-mini-its-first-llm-speeds-no-scores . The larger context window, the 100B+ model, and the specialised DevOps and cybersecurity variants are roadmap items, not releases. The Qwant deployment demonstrates the model runs in production, but the speed and cost figures come from a related-party implementation without a named baseline. OVHcloud's achievement is genuine: it has assembled a team, secured public compute, trained a novel architecture from scratch, and committed to open release with regulatory documentation. Whether the efficiency claims translate to independent deployments and whether the roadmap executes on schedule are the open questions the coming months will answer. Sources 1. OVH announces Olala mini 7A1B https://corporate.ovhcloud.com/en/newsroom/news/ovh-olala-mini corporate.ovhcloud.com 2. OVHcloud launches open AI model trained entirely on ... https://superpowerdaily.com/posts/ovhcloud-launches-open-ai-model-trained-entirely-on-european-supercomputers superpowerdaily.com 3. OVHcloud Introduces Olala Mini 7A1B as Its First Open- ... https://www.unite.ai/ovhcloud-introduces-olala-mini-7a1b-as-its-first-open-source-llm unite.ai 4. OVHcloud unveils Olala mini, its first LLM: speeds, no scores https://www.actuia.com/en/news/ovhcloud-unveils-olala-mini-its-first-llm-speeds-no-scores actuia.com 5. OVHcloud AI Summit 2026 : l’IA souveraine le 19 novembre https://blog.ovhcloud.com/fr/posts/ai-summit-2026-sovereign-ai blog.ovhcloud.com 6. https://www.hub-franceia.fr/wp-content/uploads/2025/11/FR-VF-Lancement-Dragon-15-Octobre-2025.pdf https://www.hub-franceia.fr/wp-content/uploads/2025/11/FR-VF-Lancement-Dragon-15-Octobre-2025.pdf hub-franceia.fr