OVHcloud unveils Olala Mini 7A1B, a planned open-source model
OVHcloud's Olala mini 7A1B pairs 7 billion parameters with 1 billion active, European supercomputer training and an 8 GB GPU target for local deployment.

Why it matters
- OVHcloud announced Olala mini 7A1B on November 17, 2026, with 7 billion total and 1 billion active parameters.
- OVHcloud says the model can run on an 8 GB GPU and reaches 160 tokens per second on an RTX A4500.
- OVHcloud trained the model from scratch on Jupiter, MeluXina and Jean Zay using GH200, A100 and H100 GPUs.
- OVHcloud reports that Qwant's flash answers are 20% faster, with running costs divided by three.
- Version 1.1 is planned by the end of 2026 with a 128,000-token context window.
By Elena Vasquez
OVHcloud announced Olala mini 7A1B on November 17, 2026, with 7 billion total parameters, 1 billion active for each request and a claimed ability to run on a GPU with as little as 8 GB of memory.
OVHcloud introduced Olala as its first family of large language models. The company plans to publish the model weights and a pre-trained base model on Hugging Face under the Apache 2.0 license.
OVHcloud's release package also includes compliance documentation for the AI Act, the European Union's artificial-intelligence legislation. OVHcloud's official launch presentation is scheduled for the OVHcloud AI Summit on November 19, 2026.
OVHcloud says it trained Olala mini 7A1B entirely on three European public supercomputers, rather than relying on outside labs. That choice puts public infrastructure at the center of the company's first model effort.
In OVHcloud's account, the project also assembled its AI team and led to Agentic Lab, OVHai LLM, a dedicated group for continuing the model family.
OVHcloud uses the Qwant deployment as an immediate market test, presenting it as early evidence that a smaller, internally developed model can cut response time and operating cost.
OVHcloud says Qwant has already put the model into service for flash answers. The company reports answers 20% faster and running costs divided by three against the model Qwant previously used.
CEO Octave Klaba said OVHcloud needs to create models for its own internal requirements as well as customer needs. “We cannot simply rely on the generosity of certain labs to share their LLMs as open source in the years to come,” he said.
What can Olala mini 7A1B do?
OVHcloud describes Olala mini 7A1B as a compact model that uses only part of its network for each request. Its specifications explain why the company can target modest hardware while keeping a 7-billion-parameter design.
- OVHcloud reports 7 billion total parameters in the Olala mini 7A1B architecture.
- OVHcloud says 1 billion parameters are active for each request.
- OVHcloud distributes those active parameters across 256 experts per layer, so each request uses a fraction of the full network.
- OVHcloud sets the current context window at 32,000 tokens, the amount of text the model can process in one sequence.
OVHcloud says the hybrid design combines Differential-TPA with Mamba-3, uses the BF16 number format and should reduce memory demand while increasing response speed for small deployments.
OVHcloud reports throughput of 160 tokens per second on an RTX A4500 and 50 tokens per second on a standard processor without a GPU.
The company presents those figures as evidence that its architecture can serve smaller deployments and processor-only systems. Version 1.1 will test whether the same efficiency survives a much longer context window.
What does European training infrastructure change?
OVHcloud trained the model from scratch on Jupiter, MeluXina and Jean Zay, using a mixed infrastructure of GH200, A100 and H100 graphics processors across the three European public systems.
OVHcloud describes the public support behind the project, equivalent to €7 million in computing hours, as decisive. The company says the three systems trained Olala mini without outside laboratory resources.
The initiative also assembled OVHcloud's AI team and led to Agentic Lab, OVHai LLM. OVHcloud presents the lab as an organizational foundation for later models rather than a one-off training exercise.
That structure gives OVHcloud a continuing internal path, while the public training run supplies a practical test of the economics Klaba later cites.
What has Qwant measured?
Qwant has already adopted Olala mini for its flash-answer feature. According to OVHcloud, the service now produces answers 20% faster and divides running costs by three compared with the model Qwant used before.
OVHcloud calls the Qwant result an early operational example of the deployment economics behind Olala. The company uses the deployment to support its case for compact models in production.
Those figures give the announcement a production measure rather than a laboratory claim alone.
Where does Olala go after its first release?
- OVHcloud plans version 1.1 by the end of 2026, with a 128,000-token context window and improved mathematical reasoning.
- OVHcloud plans version 2.0 to exceed 100 billion parameters, with specialized systems for business agentic AI, cloud DevOps and cybersecurity.
OVHcloud says OVHai LLM will develop the specialized models. Its cybersecurity work will include both defensive and offensive uses, while the business line targets agentic AI.
OVHcloud intends later versions to move Olala into its products and operational tools. The company expects Olala eventually to power OVHai Workspace.
OVHcloud says the encrypted version of OVHai Workspace will require models capable of running locally on a workstation. That requirement puts a hard efficiency test on future Olala releases.
OVHcloud's DevOps work includes versions for developers, architects and operations teams responsible for round-the-clock service continuity. The roadmap thus turns a compact model into a layer for products, cloud operations and security.
Can training become a business?
Klaba said training expertise has become widely accessible and that training costs fell by a factor of 20 in three years, which he presented as a route to profitable model development.
OVHcloud's open-source plan gives that argument a practical form. It plans to publish weights under Apache 2.0 and include AI Act compliance documents, pairing model access with regulatory paperwork.
OVHcloud's Olala rollout gives the company both a compact model and an in-house development path. Klaba's position is that OVHcloud cannot depend on outside labs continuing to share open models.
What has to work next?
OVHcloud's own test is whether OVHai LLM can preserve the reported efficiency and deployment gains as context expands and the family grows toward specialized systems far larger than the first release.
Version 1.1, due by the end of 2026, is the first checkpoint. It must extend context from 32,000 to 128,000 tokens while improving mathematical reasoning. OVHcloud will also need to show that the Qwant economics survive the change.
The outcome will show whether Olala can turn training on public supercomputers into profitable, specialized AI products.
Source: unite.ai
More from Elena Vasquez
Show full bio
Market editor covering technology and product news in artificial intelligence and machine learning at AI In Context.
708 articles
Related articles
- OpenAI Launches o3 and o4-mini, Its Smartest Models Yet
- OpenAI details codex-1: an o3 variant tuned for real coding work
- OpenAI Ships o1 to Developers With 60% Cheaper Audio
- OpenAI's Jalapeño Chip: LLMs Cut Design Time to Record Lows
- Anthropic and OpenAI Ship New Models With the Same Pitch: More for Less