Europe’s biggest cloud provider just became a model maker. On November 17, 2026, OVHcloud announced Olala, its first family of large language models, starting with a compact model called Olala mini 7A1B. The headline number is the one in the name: the model has 7 billion parameters in total, but only 1 billion of them are active for any single request. That is the whole pitch — a model small enough to run on ordinary hardware, built and trained in Europe, and released as open source. The official launch takes place at the OVHcloud AI Summit on November 19, where the company will present the model in detail.
The timing matters. Most of this year’s open-model news has been about scale — bigger models, bigger clusters, bigger budgets. Olala goes the other way. It is aimed at the growing number of people and companies who want an AI model they can actually run themselves, on a machine they own, without renting a data-centre GPU by the hour. If it works as advertised, it is a genuinely useful addition to the small-model wave that has been gathering all year.
How the Architecture Works

Olala mini is a mixture-of-experts model, which is how it keeps its active size so small. The 7 billion parameters are spread across 256 experts per layer, and the model routes each request to only the experts it needs — about 1 billion parameters’ worth. The rest stay idle. That is the same trick behind some of this year’s most efficient models, and it is why a 7B model can behave, in cost terms, more like a 1B one.
The attention mechanism is a hybrid: Differential-TPA combined with Mamba-3, in BF16 format. Mamba-style state-space layers are cheaper than pure attention over long sequences, which is part of how the model keeps its memory footprint down. The practical result, according to OVHcloud’s official announcement, is a model that runs on a GPU with as little as 8 GB of memory, reaching 160 tokens per second on an RTX A4500 — and, without a GPU at all, about 50 tokens per second on a standard processor. Its context window currently reaches 32,000 tokens.
That puts it in interesting company. The year has already seen a run of small, cheap models aimed at everyday work — from Tencent’s Hy4 Flash, priced for high-volume use, to Claude Haiku 5.5, Anthropic’s fastest and cheapest model yet. Olala’s answer to that trend is different: instead of a cheap API, it offers a model you can download and own.
Trained on Europe’s Public Supercomputers
Olala mini was trained from scratch on three European public supercomputers — Jupiter in Germany, MeluXina in Luxembourg and Jean Zay in France — equipped with GH200, A100 and H100 GPUs. OVHcloud says the public support was worth the equivalent of €7 million in computing hours, and the company was explicit in thanking the public authorities behind it. The project also gave rise to OVH’s Agentic Lab, OVHai LLM, the team behind the model.
That public backing is part of the story, not just background. European governments have spent years talking about digital sovereignty — the idea that Europe should be able to run its own AI infrastructure rather than renting it from American or Chinese labs. A model trained on European public machines, released under the Apache 2.0 licence, is that idea in concrete form. It is the same impulse behind projects like du’s National HyperAI in the UAE, where sovereign AI is treated as national infrastructure rather than a product choice.
The release package is also more complete than a weights dump. According to OVHcloud, Olala mini ships with its weights, its pre-trained base model, its technical report and its AI Act compliance documentation — the paperwork the EU’s AI legislation requires. It will be available as open source on Hugging Face under the Apache 2.0 licence, via OVH’s OVHaiLLM profile on Hugging Face.
What You Can Actually Run on It

The model is built for agentic work — tasks where an AI uses tools and completes jobs in several steps, rather than just answering a single question. That is the use case OVHcloud is designing around, and it already has one production user to point to: Qwant, the French search engine, has adopted Olala mini for its flash-answer feature, with answers 20% faster and running costs divided by three compared with the model it previously used.
For individual users and small teams, the interesting number is the 8 GB GPU requirement. That is the memory in a mid-range gaming graphics card, and it means the model can run on hardware many people already own — a home PC, a small office server, a laptop with a decent GPU. With tools like llama.cpp 1.0, which recently gained a stable API for local AI, running a model like this at home is now a realistic weekend project rather than a research exercise. At 50 tokens per second on a CPU alone, it is even usable without a GPU — slow by data-centre standards, but fast enough for drafting, summarising and simple agents.
Businesses get a different benefit: data that never leaves the building. A model that runs locally means customer data, internal documents and proprietary code do not have to be sent to an outside API. That is the same reason regulated industries have been watching the small-model trend so closely all year.
What’s Next: Version 1.1 and a 100B Successor
OVHcloud has already published its roadmap. Version 1.1, expected by the end of 2026, will extend the context window from 32,000 to 128,000 tokens and improve mathematical reasoning. Version 2.0 will go much further: it will exceed 100 billion parameters and split into three specialisations — agentic AI for businesses, cloud DevOps, and cybersecurity, covering both defensive and offensive aspects. The agentic line will eventually power OVHai Workspace, whose encrypted version will need models capable of running locally, directly on the workstation.
That roadmap shows OVHcloud is not treating Olala as a one-off experiment. The company wants its own models for its own internal needs and for its customers — as CEO Octave Klaba put it in the announcement, “We cannot simply rely on the generosity of certain labs to share their LLMs as open source in the years to come.” He also noted that the cost of training has been divided by 20 in three years, which is why building a model in-house is now a viable business decision rather than a moonshot.
The Honest Caveats
There are gaps worth noting before anyone plans around Olala. The announcement includes no public benchmark scores, so there is no independent way yet to compare its quality against other models of similar size — the K2-Horizon open-model fleet announced earlier this year, for example, shipped with detailed benchmark tables. The speed figures are OVHcloud’s own measurements, with no prompt length, quantization or load conditions given. And while the weights are promised for Hugging Face under Apache 2.0, the repository was not live at the time of the announcement — Unite.AI’s coverage notes the OVHaiLLM profile did not yet list Olala on November 17. The supported languages are also not stated.
None of that makes the announcement unimportant. The formal launch on November 19 should fill in several of these blanks, and the real test will be what happens when developers can download the weights and run their own evaluations. What is already clear is the direction: small, open, locally runnable models are no longer a side project of the AI industry. A major European cloud provider has now bet its own roadmap on them — and if Qwant’s early results hold up, Olala mini could be the model that brings capable AI to hardware most of us already have.

