The name reads like a spec sheet: 7A1B, seven billion parameters, one billion activated per query. Octave Klaba unveiled Olala mini on 17 November 2026, OVHcloud's first large language model, two days before its weights go online, scheduled for Thursday 19 November on Hugging Face under Apache 2.0. The announcement gives a memory figure, two speeds and one customer. It gives no scores.
Seven billion parameters, one billion at work
Olala mini is a mixture of experts: each layer contains 256 specialised sub-networks, and a router activates only some of them for each text fragment. Hence the one billion active parameters out of seven, and the speed. The attention mechanism, by which a model links words together, is described as hybrid: it combines Differential-TPA, which compresses and filters classical attention, and Mamba-3, which reads text as a stream while keeping only a fixed-size summary.
OVH announces operation on a graphics card with at least 8 GB of memory, 160 tokens per second on an RTX A4500, 50 without a graphics card, and a context window of 32,000 tokens. The model will be released with its weights, its pre-trained base model, a technical report and AI Act compliance documentation. The official launch will take place on Thursday at the OVHcloud AI Summit, at the Carrousel du Louvre.
8 GB announced, 14 GB of weights
BF16 encodes each parameter on 16 bits, i.e. two bytes. Seven billion parameters therefore weigh 14 GB, almost double the graphics memory announced as sufficient. The gap has two possible explanations: a compressed version of the model or partial loading of experts, and the announcement leaves the question open.
The speeds invite the same reading. The 160 tokens per second are measured on an RTX A4500, a workstation card with 20 GB of memory, not 8. As for the 50 tokens per second, they are obtained "on a simple processor", without naming the processor. None of these figures measures response quality, and the announcement contains no evaluation results. Without scores, it is impossible to position Olala mini against the small open models already available, from Mistral's Ministral to Google's Gemma or Alibaba's Qwen.
An architecture from Dragon LLM, trained on public compute
The architecture does not come from Roubaix. It comes from Dragon LLM, formerly Lingua Custodia, a Paris-based SME founded in 2011 around financial translation, whose acquisition OVHcloud announced on 25 March 2026 and which it renamed OVHai LLM. Winner in June 2024 of a European Commission competition, the Large AI Grand Challenge, it had obtained 3.2 million compute hours on the Leonardo and Jupiter supercomputers, Silicon.fr reported. The competition required a model of at least 30 billion parameters. Dragon LLM had presented in October 2025 a 3.6-billion demonstrator, without benchmark results.
Olala mini was trained on three public machines: Jupiter in Germany, MeluXina in Luxembourg and Jean Zay in France. This support represents the equivalent of €7 million in compute hours, writes OVH, which thanks the public authorities without specifying under what scheme these hours were allocated. The group, with revenue of €1.145 billion in its last financial year, therefore did not finance this first training alone. It employs just under 100 people on AI, in three Paris laboratories, compared with more than 1,000 employees at Mistral, according to AFP, to which Octave Klaba says he wants to show that Europe "has its place in all layers of artificial intelligence".
Qwant, a first customer that is not a third party
The only usage reference is Qwant, which already uses the model for Réponse Flash, the summaries generated at the top of its results: responses 20% faster and operating costs divided by three, compared with a previous model that is not named. Qwant is not a customer like any other. The search engine belongs to Synfonium, 75% owned by Octave and Miroslaw Klaba and 25% by Caisse des Dépôts. This figure counts as in-house feedback.
The choice of a small model responds first to an internal need. OVHai Workspace, the messaging, storage and videoconferencing suite presented in June, must exist in an encrypted version, which will require models capable of running on the workstation: that is the use the announcement assigns to Olala. Olala mini is therefore judged less as a rival to Mistral, which Octave Klaba compared in October to an "AI services company", than as a component of OVHcloud's products, after the acquisitions of Seald, Dragon LLM and Gladia.
Thursday, three checks to make
- The scores. The technical report must position the model on public evaluations, against open models of comparable size. A speed says nothing about the quality of the answers.
- The real memory. Loading the weights on an 8 GB card and measuring remains the only way to know what the announced figure covers.
- The training data. Since 2 August 2025, the AI Act requires any provider of a general-purpose AI model, even open, to have a copyright policy and a public summary of training content. The promised documentation will say what Olala learned from.
Part of the roadmap is dated. Expected by the end of the year, version 1.1 will extend the context to 128,000 tokens and must improve in mathematical reasoning. Version 2.0, undated, will exceed 100 billion parameters according to Programmez, with models specialised in enterprise agentics, cloud DevOps and cybersecurity.
7A1B says how many parameters work. The technical report will have to say what they can do.
Our articles will then appear first in Google Top Stories.
