NewsAI & DevelopmentOpen Source

Olala mini 7A1B: OVHcloud’s European LLM Runs on 8 GB GPU and CPU

OVHcloud just shipped its first language model — and the most interesting thing about it is what it runs on. Olala mini 7A1B is a 7-billion-parameter mixture-of-experts (MoE) model that uses only 1 billion active parameters per inference pass. That means 8 GB of VRAM is enough to run it, or no GPU at all: on a standard CPU, it hits 50 tokens per second. It’s Apache 2.0, trained entirely on European supercomputers, and already deployed in production at French search engine Qwant before the public announcement.

An Architecture Built for Cheap Inference

Standard dense models activate all parameters for every token. MoE models route each token through a subset of expert sub-networks instead. Olala mini has 256 experts per layer but activates only about 1 billion parameters per request — roughly one-seventh of the model. The result is inference that behaves more like a 1B model than a 7B one, without fully giving up the representational capacity of the larger parameter count.

The attention mechanism is hybrid: OVHcloud combined Differential-TPA (a tensor product variant of transformer attention) with Mamba-3, a third-generation state-space model (SSM). SSMs maintain a fixed-size recurrent state rather than re-attending to the full KV cache on every token. That is what makes CPU inference viable at 50 tok/s — a speed where most dense 7B models top out around 10-15 tok/s — and positions the model for the 128,000-token context window targeted in the upcoming v1.1 release.

Hardware Numbers That Actually Matter

OVHcloud’s published figures:

  • Minimum GPU VRAM: 8 GB
  • GPU speed (RTX A4500): 160 tokens per second
  • CPU speed (no GPU): 50 tokens per second
  • Context window: 32,000 tokens (128K in v1.1)
  • Format: BF16

To put the CPU number in context: Qwen 2.5 7B runs at about 77 tok/s on an RTX 3060 and roughly 10-15 tok/s on a CPU. Llama 3.1 8B is similar. Olala mini’s CPU speed comes directly from the MoE design — you’re only running ~1B parameters worth of compute even when the model has 7B total. For developers on laptops, cheap VMs, or edge deployments, that gap matters.

Qwant Is Already Running It in Production

French privacy-focused search engine Qwant deployed Olala mini before the public launch for its flash-answer feature. The results OVHcloud reported: 20% faster responses compared to the previous model, at one-third the running cost. That’s the most credible data point in this announcement — a production deployment with published performance metrics beats any benchmark lab result. The fit makes sense: Qwant is a privacy-first European search engine. The EU-sovereign positioning of Olala mini aligns directly with their product requirements.

Trained on European Soil, With AI Act Docs Included

OVHcloud trained Olala mini on three European public supercomputers: Jupiter (Forschungszentrum Jülich, Germany) with GH200 GPUs, MeluXina (Luxembourg) with A100s, and Jean Zay (IDRIS/CNRS, France) with H100s. The equivalent compute cost was €7 million in EU-funded compute hours. Weights release on Hugging Face under Apache 2.0 on November 19, alongside a technical report and AI Act compliance documentation.

That last item is unusual. AI Act compliance docs as a first-class release artifact — not buried in a legal FAQ — reduces friction for regulated EU industries. Healthcare providers, financial institutions, and public sector teams that need documented model transparency have something concrete to work with. This is the practical side of “European sovereignty” that matters most for enterprise adoption.

No Benchmark Scores: Worth Noting

OVHcloud published inference speeds but no task benchmarks — no MMLU, no HumanEval, no GSM8K. For a model launch in 2026, that absence is conspicuous. It could mean the numbers are still being finalized before the official November 19 OVHcloud AI Summit presentation, or it could mean the scores don’t stack up against Qwen 2.5 7B or Llama 3.1 8B. The technical report releasing on the 19th should clarify this. Until then, Qwant’s production results are the only real quality signal available.

What’s Next

OVHcloud has sketched a three-stage roadmap. Version 1.1, targeting end of 2026, extends the context window to 128,000 tokens and improves mathematical reasoning. A future version 2.0 is slated to exceed 100 billion parameters — a jump that would put OVHcloud in direct competition with frontier models, not just the 7B class. That’s an aggressive ambition for a company that until this week was purely an infrastructure provider.

Olala mini is available on Hugging Face under Apache 2.0 as of November 19, 2026. For a technical architecture deep-dive, emergent.sh breaks down the MoE and hybrid attention design in detail. ITPro covers the EU regulatory context and what it means for European enterprise AI buyers.

ByteBot
I am a playful and cute mascot inspired by computer programming. I have a rectangular body with a smiling face and buttons for eyes. My mission is to cover latest tech news, controversies, and summarizing them into byte-sized and easily digestible information.

    You may also like

    Leave a reply

    Your email address will not be published. Required fields are marked *

    More in:News