Innovation & Tech

OVHcloud Ships Olala Mini 7A1B, Its First Open-Source LLM

OVHcloud's first LLM, Olala mini 7A1B, runs on 8 GB GPUs, ships under Apache 2.0, and cut Qwant's answer costs by three while training on €7 million of European supercomputing.

By Grace Kim

3 min read

OVHcloud Introduces Olala Mini 7A1B as Its First Open-Source LLM
OVHcloud Introduces Olala Mini 7A1B as Its First Open-Source LLMAI-generated

What's News

  • OVHcloud announced Olala mini 7A1B on November 17, 2026, its first open-source LLM under Apache 2.0 on Hugging Face.
  • The 7-billion-parameter model activates 1 billion parameters per request and runs on GPUs with 8 GB of memory.
  • Training used three European public supercomputers backed by computing hours worth €7 million.
  • Qwant reports 20% faster flash answers and running costs divided by three after adopting the model.
  • Version 2.0 is planned to exceed 100 billion parameters, with a 128,000-token context window due in version 1.1 by end-2026.

OVHcloud announced Olala, its first family of large language models, on November 17, 2026, opening with a compact model that runs on a GPU with as little as 8 GB of memory. The company plans to release Olala mini 7A1B as open source on Hugging Face under the Apache 2.0 licence, and it will present the model officially at the OVHcloud AI Summit on November 19, 2026.

The release package includes the model weights, a pre-trained base model and compliance documentation for the AI Act, the European Union's artificial-intelligence legislation. That documentation matters for European buyers who need regulatory cover, not just weights on a hub.

How fast is Olala mini?

The "7A1B" name encodes the architecture: 7 billion parameters total, with only 1 billion active per request. Those active parameters spread across 256 experts per layer, so each request touches a fraction of the network instead of the whole thing. The model uses a hybrid attention design combining Differential-TPA and Mamba-3, and runs in BF16 format.

OVHcloud reports throughput of 160 tokens per second on an RTX A4500 and 50 tokens per second on a standard processor with no GPU at all. The context window sits at 32,000 tokens for this first version. The takeaway: the design targets hardware that most businesses already own.

Where was it trained?

OVHcloud trained Olala mini from scratch on three European public supercomputers: Jupiter, MeluXina and Jean Zay, using GH200, A100 and H100 GPUs. The company describes the public support behind the project — equivalent to €7 million in computing hours — as decisive.

The project also did organizational work. It assembled OVHcloud's AI team and produced the Agentic Lab, branded OVHai LLM, a dedicated group that will keep developing the family rather than treat Olala mini as a one-off training run.

Who is already using it?

Qwant has adopted Olala mini for its flash-answer feature. OVHcloud reports answers arriving 20% faster, with running costs divided by three compared with the model Qwant previously used. That's the clearest early proof point for the deployment economics OVHcloud is chasing — same job, one-third the bill.

What comes next?

The roadmap calls for version 1.1 by the end of 2026, extending the context window to 128,000 tokens and improving mathematical reasoning. Version 2.0 is planned to exceed 100 billion parameters, with OVHai LLM building specialized models for business agentic AI, cloud DevOps and cybersecurity — including both defensive and offensive uses.

Later versions are meant to move the family into OVHcloud's own products and operational tools. Olala is eventually expected to power OVHai Workspace, whose encrypted version will require models that run locally on a workstation. The DevOps track covers versions for developers, architects and operations teams responsible for round-the-clock service continuity.

Chief executive Octave Klaba framed the strategy as self-reliance. OVHcloud needs models for its own internal requirements as well as customer needs, he said, adding: "We cannot simply rely on the generosity of certain labs to share their LLMs as open source in the years to come." Klaba said training expertise has become widely accessible and costs have fallen by a factor of 20 in three years — changes he presents as a route to profitable model development.

The open question is execution. OVHai LLM must carry the efficiency gains reported for a 7A1B model into the 128,000-token release and the 100-billion-plus specialized systems on the roadmap.

Source: unite.ai

Share this article:

More from Grace Kim

Grace Kim

Show full bio

Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.

2,213 articles

Related articles

« Previous articleNext article »