Back to News
ProductAI Understanding briefing

OVHcloud releases Olaka mini 7A1B, its first open‑source 7‑billion‑parameter LLM

French cloud provider OVHcloud unveiled Olaka mini 7A1B, a 7‑billion‑parameter open‑source language model that runs on 8 GB GPUs or CPUs, and is already being used by Qwant for faster search answers.

4 min readRead the linked source
Source-provided image accompanying OVHcloud releases Olaka mini 7A1B, its first open‑source 7‑billion‑parameter LLM
Source referenceSource recorded
Publisher
itbrief.co.uk
Source type
Linked source — primary-source status has not been established.
ContextUnderstand this in 60 seconds

Key terms

Large Language Model (LLM)
A language model trained on massive text corpora to generate and analyze text.
Parameter
A learned weight inside a model that influences its outputs.
Memory (Agent Memory)
Stored context an AI agent uses across steps or sessions to improve continuity.

What happened

OVHcloud announced the launch of Olaka, its first family of large language models, with the initial offering called Olaka mini 7A1B. The model is a 7‑billion‑ LLM built on a mixture‑of‑experts architecture that activates only a fraction of its parameters per request. It was trained on three European public supercomputers—Jupiter, MeluXina and Jean Zay—and is released under an Apache 2.0 licence on Hugging Face together with weights, a technical report and AI‑Act compliance documentation. Early adopters include French search engine Qwant, which reports a 20 % speed improvement and a three‑fold reduction in running costs compared with its previous model. OVHcloud says later Olaka versions will target business agents, cloud operations and cybersecurity, and a future model will exceed 100 billion parameters.

OVHcloud introduced the Olaka family of large language models, starting with Olaka mini 7A1B, a 7‑billion‑ LLM built on a mixture‑of‑experts design that activates only one billion parameters per inference step. This architecture reduces memory usage, allowing the model to run on modest hardware.

Training was performed on three European public supercomputers—Jupiter, MeluXina and Jean Zay—highlighting a fully European compute pipeline. The company released the model weights, a pre‑trained base, a technical report and AI‑Act compliance documentation on Hugging Face under an Apache 2.0 licence.

Performance claims include 160 tokens per second on an RTX A4500 GPU and 50 tokens per second on a standard CPU, with a 32,000‑token context window. OVHcloud states the model can operate on GPUs with as little as 8 GB of memory or even on CPUs without a GPU.

Qwant, a French search engine, is an early adopter. According to OVHcloud, Qwant saw answer generation 20 % faster and a three‑fold reduction in running costs compared with its previous model.

OVHcloud’s CEO Octave Klaba emphasized the strategic importance of owning model‑training capabilities, noting that training costs have fallen dramatically and that the company aims to make model development profitable. Future Olaka versions will target specific business domains and eventually exceed 100 billion parameters.

Source details: itbrief.co.uk

Why it matters

The launch marks a concrete step for European AI independence by showing that a regional cloud provider can train and open‑source a competitive LLM without relying on US‑based compute or model ecosystems. The model’s low hardware requirements—running on a GPU with as little as 8 GB of memory or even on a CPU—lower the barrier for organisations that lack large‑scale GPU clusters, potentially expanding local deployment of generative AI for privacy‑sensitive or cost‑constrained use cases. By publishing the model, weights and compliance documentation, OVHcloud gives developers transparency and the ability to fine‑tune or audit the model, addressing growing regulatory and data‑sovereignty concerns in Europe. The partnership with Qwant demonstrates immediate commercial benefit, suggesting that other European firms could adopt Olaka to reduce latency and operating expenses. If OVHcloud’s roadmap to a 100‑billion‑ model materialises, it could further challenge the dominance of US‑origin models and stimulate a more diversified AI market.

European AI sovereignty: By training and open‑sourcing a model entirely on European infrastructure, OVHcloud demonstrates a path toward reduced reliance on US‑based cloud and model providers, aligning with EU policy goals around data sovereignty and AI regulation.

Lower hardware barrier: The ability to run a 7‑billion‑ LLM on 8 GB GPUs or CPUs makes generative AI accessible to smaller firms, research labs, and public sector organisations that cannot afford large GPU clusters, potentially widening AI adoption across the continent.

Transparency and compliance: Publishing the model under an Apache 2.0 licence, together with AI‑Act compliance documentation, gives users the ability to audit, fine‑tune, and certify the model for regulated environments, addressing growing concerns about black‑box AI systems.

Commercial validation: Qwant’s reported speed and cost improvements provide early evidence that the model can deliver tangible business value, encouraging other enterprises to evaluate Olaka for similar use cases such as search, recommendation, or internal knowledge bases.

Strategic diversification for OVHcloud: Adding a software layer to its cloud services portfolio positions OVHcloud as more than an infrastructure provider, potentially opening new revenue streams through model licensing, managed services, and custom AI solutions.

Interactive Mechanism

Interactive Mechanism: How It Actually Works

Explore the underlying technology behind this development interactively.

Agent Lifecycle Stage:
1
User Intent & Planning: "Audit customer refund request #4092 and settle payment."
2
Tool Calling: Emits structured JSON call crm_get_transaction(id='4092').
3
Guardrail & Verification:Paused: High-value action requires human operator sign-off.
4
Final Settlement: Refund recorded, email receipt dispatched, and audit log stored.
Core takeaway: An AI agent is not just a language model—it is a closed loop of planning, tool invocation, and environment feedback. Production systems require self-healing retries and strict human approval guardrails.
Interactive Concept Check+10 Points
AI Models Explained Quiz

Which component of an AI application is the machine-learning model itself?

What to watch next

Key indicators to monitor include adoption rates among European enterprises, especially those with strict data‑localisation rules, and the performance and cost metrics reported by early users beyond Qwant. Watch for the release of larger Olaka models and any announced partnerships with other cloud or software providers. Regulatory response to the AI‑Act compliance documentation could also influence how other firms evaluate open‑source models for compliance. Finally, observe whether OVHcloud’s model‑as‑a‑service offerings generate sufficient revenue to sustain further large‑scale model training in Europe.

Adoption metrics: Track the number of organisations downloading and deploying Olaka from Hugging Face, especially those in regulated sectors like finance, healthcare, and government.

Performance validation: Independent benchmarks comparing Olaka’s latency, token throughput, and cost against other open‑source models (e.g., LLaMA, Mistral) will clarify its competitive standing.

Regulatory impact: Observe how EU regulators respond to OVHcloud’s AI‑Act compliance documentation and whether it becomes a reference point for future open‑source model disclosures.

Roadmap execution: The release of larger Olaka models, particularly a 100‑billion‑ version, will test OVHcloud’s claim that European compute resources can support training at that scale.

Ecosystem partnerships: Announcements of collaborations with other European cloud providers, software vendors, or academic institutions could amplify the model’s reach and influence.

Related guides & quizzes

Found this useful?