HomeNews

OVHcloud Olala Mini: 7B MoE Model with 1B Active Params

Samraat
Samraat
•
Nov 17, 2026 5:13 PM
•
0
 min read
Select Emergent as your Preferred news source
OVHcloud Olala Mini: 7B MoE Model with 1B Active Params

💡 TL;DR

  • OVHcloud released Olala Mini, a 7B MoE model with 1B active parameters and 256 experts per layer under Apache 2.0 license.
  • The model delivers 160 tokens per second on RTX A4500 GPU and 50 tokens per second on CPU-only systems.
  • Qwant already deployed Olala Mini for flash answers, achieving 20% faster responses and cutting costs by two-thirds.

OVHcloud officially released Olala Mini on November 17, 2026, marking the European cloud provider's entry into large language model development. The 7-billion-parameter mixture-of-experts architecture activates just 1 billion parameters per inference request, spreading computation across 256 experts per layer. Released under Apache 2.0 license with full weights, base model, and EU AI Act compliance documentation on Hugging Face, the model represents a strategic push for European AI sovereignty.

Hybrid Architecture Enables Efficient Inference

Olala Mini combines Differential-TPA and Mamba-3 attention mechanisms in a novel hybrid design. The sparse activation pattern keeps only a fraction of the 7B total parameters active for each query, delivering 160 tokens per second on an RTX A4500 GPU with 8GB memory. The model maintains 50 tokens per second on CPU-only systems, making it accessible for deployment without dedicated accelerators. The current release supports a 32,000-token context window in BF16 format.

Trained on European Public Supercomputers

OVHcloud trained Olala Mini across three European HPC facilities: Jupiter (GH200 GPUs), MeluXina (A100 GPUs), and Jean Zay (H100 GPUs). The public compute allocation, valued at approximately €7 million, enabled the development of OVHcloud's Agentic Lab and OVHai LLM team. This partnership demonstrates viable pathways for European organizations to develop competitive language models using shared research infrastructure rather than relying exclusively on hyperscaler resources.

Release Date and Early Production Use

Officially launched on November 17, 2026, Olala Mini entered production immediately at Qwant, the European search engine. Qwant deployed the model for its flash-answer feature, reporting 20% faster response generation and a 66% reduction in operating costs compared to the previous model. The full technical unveiling occurred at OVHcloud's AI Summit on November 19, 2026, where the company detailed its broader cloud AI strategy.

Roadmap for Extended Context and Scale

OVHcloud plans version 1.1 for late 2026, expanding the context window to 128,000 tokens and improving mathematical reasoning capabilities. Version 2.0, slated for 2027, will exceed 100 billion parameters and target three verticals: agentic AI for OVHai Workspace, cloud DevOps tooling for developers and operations teams, and cybersecurity applications covering both defensive and offensive use cases. CEO Octave Klaba emphasized that training costs have dropped by 95% over three years, making in-house model development economically viable for cloud providers.

What This Means

Olala Mini signals that mid-tier cloud providers can now train production-grade language models using public compute grants and efficient architectures. The 7B-to-1B sparse activation ratio demonstrates that mixture-of-experts topologies can deliver practical inference speeds on consumer hardware, expanding deployment options beyond datacenter GPUs. For European organizations prioritizing data sovereignty and AI Act compliance, OVHcloud's release provides a fully documented, Apache-licensed alternative to U.S. hyperscaler models.

About the writer

Samraat is a Software Engineer at Emergent, contributing to the platform's AI-powered app building systems. He previously built products at Juspay, and holds a Computer Science degree from UIET Panjab University.

Start Building
on Emergent today
Try Emergent