OVHcloud releases open 7B model that runs on 8GB GPU
OVHcloud unveiled Olala mini 7A1B, a 7-billion-parameter open-weight model under Apache 2.0.

· Last verified: 18 Nov 2026 (OVHcloud Newsroom)

Summary
- OVHcloud unveiled Olala mini 7A1B, a 7-billion-parameter open-weight model under Apache 2.0.
- It was trained on three public European supercomputers with roughly 7 million euros in granted compute hours.
- Enterprises get open weights plus EU AI Act compliance documentation, lowering the cost of self-hosted AI.
The latest
A 7-billion-parameter language model that runs on a graphics card with just 8GB of memory is now open to anyone under an Apache 2.0 license. OVHcloud announced Olala mini 7A1B on 17 November 2026, the first entry in its Olala family, with a formal launch set for the OVHcloud AI Summit on 19 November. Weights are published on Hugging Face.
Details
- The architecture: Olala mini 7A1B carries 7 billion total parameters but activates roughly 1 billion per request, spread across 256 experts in each layer, according to OVHcloud. It uses a hybrid attention design combining Differential-TPA and Mamba-3, in BF16 format, with a context window currently set at 32,000 tokens.
- The hardware claim: OVHcloud said the model runs on a GPU with only 8GB of memory, reaching about 160 tokens per second on an RTX A4500. On an ordinary CPU with no graphics card, it still generates roughly 50 tokens per second, putting local deployment within reach of standard office machines.
- European training base: The model was trained on three publicly funded European supercomputers — Jupiter, MeluXina and Jean Zay — using GH200, A100 and H100 processors. OVHcloud valued the public compute support at the equivalent of 7 million euros in machine hours. The company did not specify how many tokens the model was trained on.
- What ships with it: The release includes the open weights, the pre-trained base model, a technical report, and compliance documentation for the European Union's AI Act, aimed at enterprises that must demonstrate regulatory conformity before production. The announcement did not specify which languages the model supports.
- Already in production: OVHcloud said the model is running at the Qwant search engine, powering its quick-answers feature. The company reported the switch made that feature 20% faster and cut operating costs to one third of what the previous model required.
- The CEO's case: Octave Klaba, chief executive, argued that artificial intelligence is strategic for OVHcloud and that the company needs to build its own models rather than depend on other laboratories releasing theirs as open source. He said training expertise is now widely available and training costs have fallen 20-fold in three years.
- The roadmap: OVHcloud expects a 1.1 release by the end of 2026, extending the context window to 128,000 tokens and improving mathematical reasoning, with no precise date given. Version 2.0 is planned to exceed 100 billion parameters, built around specialised models for enterprise agentic AI.
- Version 2.0 targets: The planned products include an encrypted OVHai Workspace intended to run locally, tooling for cloud development operations aimed at developers, engineers and operations teams, and both defensive and offensive cybersecurity applications.
Background
OVHcloud is a European cloud provider that has positioned itself against American hyperscalers on data sovereignty. Olala mini 7A1B is its first in-house language model family, built on a new architecture rather than fine-tuned from an existing open model.
Between the lines
The combination on display is less about raw capability than about removing the obstacles enterprises cite most. An Apache 2.0 license removes legal friction, AI Act documentation removes regulatory friction, and an 8GB memory requirement removes infrastructure cost. Training on publicly funded European machines reinforces the sovereignty pitch Klaba made, and the Qwant deployment supplies a working cost figure rather than a benchmark score.
What's next
The formal launch at the OVHcloud AI Summit on 19 November, and the version 1.1 release promised by the end of 2026 with a 128,000-token context window and stronger mathematical reasoning.
Source: OVHcloud Newsroom