OVHcloud Unveils Olala Mini, an Open-Source European LLM That Runs on 8 GB of GPU Memory

OVHcloud announced Olala, its first family of LLMs, on 17 November 2026. Olala mini 7A1B is a 7B-parameter mixture-of-experts model with 1B active parameters, released as open source under Apache 2.0 and able to run on an 8 GB GPU.

Christopher Ort
By Christopher Ort · 2 min read
Share
A graphics card and small server tower on a workbench in a data center aisle.
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. Background
  5. How it works
  6. Why it matters
  7. What's still open
  8. Sources
  9. i10x Perspective

OVHcloud has announced Olala, its first family of large language models, starting with Olala mini 7A1B, a compact open-source model trained from scratch. The French cloud provider announced it in Roubaix on 17 November 2026, ahead of an official launch at the OVHcloud AI Summit on 19 November.

i10X Quick Take

Quick Take

  1. Model: Olala mini 7A1B, a mixture-of-experts model with 7 billion parameters in total and 1 billion active, spread across 256 experts per layer.
  2. Architecture: hybrid attention combining Differential-TPA and Mamba-3, in BF16 format.
  3. Speed: runs on a GPU with as little as 8 GB of memory, reaching 160 tokens per second on an RTX A4500, and 50 tokens per second on a standard CPU. Context window is 32,000 tokens.
  4. License: open source under Apache 2.0 on Hugging Face, with weights, the pre-trained base model, a technical report and AI Act compliance documentation.
  5. Roadmap: version 1.1 by the end of the year with a 128,000 token context window; version 2.0 will exceed 100 billion parameters.

In plain English

OVH, a European cloud company, built its own small AI model instead of reselling someone else’s. Because only a fraction of the model works on each request, it is fast and cheap enough to run on a modest graphics card or even a laptop-class processor. It is released openly, so anyone can download and use it.

What happened

OVH said Olala mini shows “that European companies can innovate and train agentic LLMs from scratch.” Qwant has already adopted the model for its flash answer feature, which OVH says is 20% faster and costs one third as much to run compared with the model Qwant used before.

Background

The model was trained on three European public supercomputers, Jupiter, MeluXina and Jean Zay, using GH200, A100 and H100 GPUs. OVH puts the public support at about 7 million euros in computing hours. The effort gave rise to OVH’s Agentic Lab, OVHai LLM.

How it works

A mixture-of-experts design activates only a small share of its parameters for each request, which cuts memory use and speeds up answers. The hybrid attention mixes a differential attention variant with Mamba-3, a state-space approach, to keep long inputs efficient. OVH plans later versions for three areas: agentic AI for business, cloud DevOps and cybersecurity.

Octave Klaba, CEO of OVHcloud, said: “We cannot simply rely on the generosity of certain labs to share their LLMs as open source in the years to come.”

Why it matters

For European companies, it is a locally trained, openly licensed model with AI Act documentation included, which helps with compliance and sovereignty requirements.

For developers, a model that runs on 8 GB of GPU memory lowers the cost of local agent experiments.

For the market, a cloud provider training its own models adds another European option next to larger labs.

What’s still open

  • Independent benchmark results; OVH has not published comparisons in its announcement.
  • How well a 1 billion active parameter model handles agentic tasks beyond Qwant’s flash answers.
  • Details at the official launch on 19 November.

Sources

  • OVH, “OVH announces Olala mini 7A1B,” OVHcloud newsroom, 17 November 2026

i10x Perspective

Small, efficient open models matter most where privacy and cost rule out frontier APIs. OVH’s pitch is as much about control as capability. For the wider context, read our piece on European AI sovereignty and browse the Models section. Another open-weight release this week is the Nemotron for Science launch.

Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

ModelsNews

Editor's picks

  1. Genesys Launches Standalone AI Agents and Arena, With Pricing Based on How Hard Each Task Is
  2. Seedance 2.5 4K Reaches Developers: 10-Bit Video up to 30 Seconds With 480p Draft Mode
  3. Perplexity Computer Adds Multiagent Mode, Where Up to 16 Agents on Rival Models Check Each Other
  4. Gemini Can Now Make Calls, Pay Bills and Shop, Free in Six Countries With Gemini 4 Hydrogen
  5. Hinton on Firing Line: AI Is Making ‘Alien Beings’ and Self-Regulation Is ‘Absurd’