The name reads like a technical sheet: 7A1B, seven billion parameters, one billion of which are active at a time. Octave Klaba unveiled Olala mini, OVHcloud’s first major language model, on November 17, 2026, two days before the upload of its weights, scheduled for Thursday November 19 on Hugging Face under the Apache 2.0 license. The announcement gives one memory, two speeds and one client. It gives no score.
Seven billion parameters, a billion at work
Olala mini is a mixture of experts: each layer contains 256 specialized subnetworks, and a switch only uses a part of them for each text fragment. Hence the billion active parameters out of seven, and the speed. OVH talks about hybrid attention. Differential-TPA is the attention component, the mechanism by which a model links words together: its name combines two published techniques, one which filters noise, the other which reduces the memory consumed. Mamba-3 belongs to another family, state space models, which read text as a stream and only keep a state of fixed size.
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, pre-trained base model, 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 Go annoncés, 14 Go en BF16
The BF16 format encodes each parameter in 16 bits, or two bytes. At this precision, seven billion parameters weigh 14 GB, almost double the graphics memory announced as sufficient. The likely explanation is banal: a quantized version, where each parameter is stored on fewer bits. At 8 bits, the weights drop to 7 GB, and OVHai LLM is already releasing other models in a variation of this type. It remains to be seen which version the 8 GB corresponds to, and whether the speeds have been measured with it.
The flow rates call for the same reading. The 160 tokens per second are measured on an RTX A4500, a workstation card which has 20 GB of memory, not 8. As for the 50 tokens per second, they are obtained “on a simple processor”, without the processor being named. None of these numbers measure the quality of responses. Without a score, it is impossible to compare Olala mini to the small open models already available, from Mistral’s Ministral to Google’s Gemma or Alibaba’s Qwen. A European point of comparison has existed since November 12: the Italian Translated published Lara Mini Base, an expert mix of 7.2 billion parameters including 0.6 active, with its scores but under a non-commercial license, when the Apache 2.0 license promised for Olala authorizes commercial use.
Behind Olala, the Dragon LLM and public computing team
The announcement names OVHai LLM, the new name of Dragon LLM, the former Lingua Custodia, a Parisian SME founded in 2011 around financial translation, of which OVHcloud announced the acquisition on March 25, 2026. This team was already working on a hybrid architecture of the same family. Winner in June 2024 of a European Commission competition, the Large AI Grand Challenge, she obtained 3.2 million hours of calculation on the Leonardo and Jupiter supercomputers, reported Silicon.fr. Dragon LLM presented in October 2025 a demonstrator with 3.6 billion parameters. Its Hugging Face record displays an average of 73.27% over seven tests, tied with SmolLM3 and ahead of Qwen3-4B (71.56%). The team therefore knows how to publish scores.
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 euros of computing hours, writes OVH, which thanks the public authorities without specifying in what capacity these hours were allocated. The group, with a turnover of 1.145 billion euros in its last financial year, therefore did not finance this first training alone. It employs a little less than 100 people on AI, in three Parisian laboratories, compared to more than 1,000 employees at Mistral, according to AFP, to whom Octave Klaba said he wanted to show that Europe “has its place in all layers of artificial intelligence”.
Qwant, a first customer who is not a third party
The only reference in use is Qwant, which already uses the model for Flash Response, the summaries generated at the top of its results: responses 20% faster and operating costs divided by three, compared to a previous model that OVH does not name. On October 9, Next wrote that these flash responses were powered by Mistral models, and Qwant told him: “Mistral’s recent pricing developments lead us to evaluate the available alternatives. » 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. These two figures therefore come from a Klaba family company.
The choice of a small model first responds 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: this is the use that the announcement assigns to Olala. Olala mini is therefore seen first and foremost as a component of OVHcloud products, after the acquisitions of Seald, Dragon LLM and Gladia. If the model replaced at Qwant is indeed that of Mistral, its first effect will have been to remove a client from the company that Octave Klaba compared in October to an “AI ESN”.
Thursday, three checks to do
- The scores. The technical report must place the model on public evaluations, compared to open models of comparable size. A flow rate says nothing about what the answers are worth.
- Real memory. Loading the weights onto an 8 GB card and measuring remains the only way to know what the announced figure covers. The team’s previous model, Dragon 3B, requires specific code supplied with the weights: a CIO will want to know if Olala runs in the engines it already uses, llama.cpp, Ollama or vLLM.
- Training data. Since August 2, 2025, the AI Act requires any provider of general-purpose AI models, even open ones, to have a copyright policy and a public summary of training content. The promised documentation will say what Olala learned about.
Expected by the end of the year, version 1.1 will bring the context to 128,000 tokens and must progress in mathematical reasoning. Version 2.0, undated, will exceed 100 billion parameters according to Programmez, with models specialized in enterprise agentics, cloud DevOps and cybersecurity.
7A1B says how many parameters work. The technical report should say what they can do.
Our articles will be prioritized in Google’s “Featured” section.

