Skip to content

Sovereign AI / Olala evaluation

OVHai and Olala for industrial and enterprise systems: test the claims, own the decision.

Decide whether Olala and the wider OVHai catalogue fit your workload. Turn provider announcements into measured evidence, deployment controls and the next agreed milestone. Hyperion is independent of OVH and OVHcloud.

Independent product leadership. Worldwide, with an EU focus.

Discuss your projectExplore the decisions
Independent product architectureChoose the components. Own the decision.
Your workload & acceptance criteria
Qualify each layer
  1. Application & authorityYour workflow · your approvals and existing tools
    01
  2. Model & adaptationQualified release · configuration · retrieval · justified tuning
    02
  3. Execution & operationSuitable self-managed runtime or contracted hosted services
    03

Inspect the connections.Model execution ≠ workflow history ≠ tool access

Evaluate → accept, revise or replace

Application, model and runtime choices for the selected workload.

Begin with a useful application.

A compact model is worth evaluating where size, cost or connectivity constrains the workload. Establish a useful workflow, its operating owner and the evidence required for acceptance. These examples illustrate possible scopes; they are not client case studies.

Offline and edge assistance

A technician question or site record → bounded assistance on local hardware → a qualified person who decides the next action. OVH states that Olala mini runs on a GPU with 8 GB of memory and on a CPU without a GPU. Verify memory, sustained throughput and failure behaviour on the exact device.

Does the target device meet the workload without a network connection?

Edge qualification

High-volume text tasks

Summaries, extraction, classification or log triage → a smaller model replacing a larger one → a regression check against the current output. Measure quality, latency and operating cost on your own records before switching.

Does the smaller model hold the agreed quality at lower cost?

Applied AI systems

Engineering knowledge

Approved manuals and engineering records → a cited answer or explicit abstention → an engineer who checks its relevance. The announced 32,000-token context limits how much retrieved material fits in one request; design retrieval around it.

Can the user verify the answer against an approved source?

Private knowledge and RAG

Edge assistance does not establish autonomous machine control or a safety case. Perception, planning, actuation and independent assurance require separately defined engineering and evidence.

Choose the layer you need.

OVHai builds the models; OVHcloud operates the infrastructure and hosted services. They are different products with different terms. A project does not need every layer; qualify access, licence, deployment option and supplier obligations for the ones it uses.

Olala models

OVH's own model family, announced on 17 November 2026. Olala mini 7A1B is a mixture-of-experts model with 7 billion parameters in total and about 1 billion active, announced under Apache 2.0 with weights, the pre-trained base model, a technical report and AI Act documentation.

OVH announcement (French)

OVHai specialised models

The OVHai LLM catalogue on Hugging Face also lists finance models and a Dragon 3B base model. The repositories checked are gated and licences differ: some are Apache 2.0, others follow the Llama 3.1 licence of their base model. Check each one.

OVHai LLM on Hugging Face

AI Endpoints

OVHcloud's hosted, pay-per-token inference service. Confirm that the exact model you need is listed, its lifecycle and retirement notice, and the data handling that applies to your requests.

OVHcloud AI Endpoints

Compute

GPU instances and managed deployment for self-hosted models. Translate the workload into capacity and operating requirements, then coordinate the choice with the responsible supplier. Hyperion is not the compute supplier.

OVHcloud GPU instances

OVH also announced OVHai Workspace, an agentic collaboration suite that Olala is to power. It is not assessed here until its terms and availability are published.

Prove the fit before rollout.

OVH's 17 November press release reports throughput and gives no quality benchmark. Start with a decision record and an evaluation baseline on your own tasks, compared with suitable alternatives such as Ministral, Gemma or Qwen. Choose the smallest change that satisfies the acceptance criteria.

The model record

Exact release and revision, licence and access conditions, runtime, full memory requirement on your hardware, supported languages, intended tasks, known limitations and maintenance owner.

The evaluation baseline

Your tasks, your documents and your languages, scored against the model you would otherwise use. Record quality, refusals, failure modes, latency and cost. Provider throughput figures do not measure answer quality.

The acceptance decision

Review observed quality, failures, latency and operating cost against agreed constraints. Proceed, revise, select another model or stop. A launch announcement or a reference customer alone does not establish acceptance.

OVH has announced a version 1.1 with a 128,000-token context by the end of 2026, and a version 2.0 above 100 billion parameters with specialised models for enterprise agents, cloud DevOps and cybersecurity. Keep the evaluation set so each release can be retested before adoption.

Model evaluation and adaptation in detail

Follow the complete data path.

Model execution, retrieval, workflow state and operational records can sit in different places. Qualify them separately; a European supplier or a local model does not make every part of the system local.

Self-managed weights

Olala uses a hybrid attention design (Differential-TPA with Mamba-3). Confirm that your serving runtime supports it. If a release needs repository-supplied code to run, review that code as software you execute.

Example: a custom-code OVHai release

Memory and speed claims

OVH states 160 tokens per second on an RTX A4500 and 50 tokens per second on a CPU. Seven billion parameters in BF16 occupy about 14 GB before runtime overhead; establish how the announced 8 GB minimum is reached on your hardware.

Hosted inference

AI Endpoints is billed per token and models can be retired. Confirm model availability, region, retention and the retirement notice period before building a dependency on a hosted model.

Read the billing and lifecycle terms

Identity and tools

Keep tool permissions and required confirmations outside model discretion. Test a refused action, a revoked permission and an uncertain external effect before accepting the workflow.

Security for AI and AI for security workflows

Provider material reviewed on 17 November 2026, before the Olala weights were published. Features and terms change; recheck the selected release and service at scoping and acceptance. This page does not certify a particular deployment or imply automatic regulatory compliance.

Keep the product decision on your side.

OVH offers its own services. Buy directly when the supplier already meets the requirement. Hyperion's role in an agreed engagement is independent client-side product leadership: decide whether OVHai fits, connect suppliers to one milestone and define acceptance with the people who will operate the result.

Decision and scope

A workload brief, model and deployment comparison, acceptance criteria, risks, dependencies and explicit delivery responsibilities.

Agreed implementation

Only the code, configuration, integration or adaptation work named in the written scope, with its evaluation evidence and handover. An engagement does not imply an unlimited delivery team or 24/7 operations.

Capability transfer

An operating owner, inspectable configuration, evaluation and change process, escalation and replacement plan. Connect the work to your product roadmap rather than leaving an isolated demonstration.

Hyperion is independent and is not affiliated with or endorsed by OVH or OVHcloud. This website is hosted on OVHcloud as an ordinary customer. This page does not claim OVH partnership, certification, resale rights, vendor escalation privileges or completed customer deployments.

Hyperion
The product decisions, implementation, evaluation and handover named in the engagement.
Your organisation
Data rights, system ownership, approvals, an operating owner and the required engineering and safety resources.
Technology suppliers
Their contracted models, platforms, infrastructure, integrations, support and release obligations.

Deliverable IP rights follow the engagement terms; model rights follow the applicable licence or vendor contract. Product leadership does not replace independent legal, safety or cybersecurity assurance.

Choose the responsibility you need.

From a startup's first industrial workflow to an enterprise decision across suppliers: worldwide delivery with an EU focus, scoped to actual capacity and requirements. Fees are provided by quotation.

  1. Which approach should we choose?

    Product Decision Review

    Frame the requirements, review available evidence, compare options and establish the next decision. Where evidence is missing, define the evaluation needed. Implementation is excluded.

  2. Can we establish one accepted capability?

    Product Leadership Mission

    Lead one agreed milestone, with the code, configuration, integration, evaluation and handover explicitly included in scope. Name client and specialist capacity before delivery commitments are made.

  3. Who holds the product decisions across teams?

    Executive Product Leadership

    Establish an agreed leadership mandate, priorities, decision rights and acceptance cadence across products and suppliers. Engineering, operations and independent assurance capacity remain explicitly allocated.

Discuss your project

Practical questions.

Will you always recommend OVHai?

No. The decision should follow the workload, evidence, rights, deployment constraints and operating capacity. The comparison may favour Mistral, another open-weight model, a managed service, a conventional system or no AI.

Is OVHai the same as OVHcloud?

No. OVHai LLM is OVH's model lab, built on Dragon LLM (formerly Lingua Custodia), which OVH announced it was acquiring in March 2026. OVHcloud supplies the infrastructure and hosted services, such as AI Endpoints and GPU instances. Qualify and contract each layer separately.

Can Olala run on our own hardware?

OVH announced that Olala mini runs on a GPU with at least 8 GB of memory, or on a CPU without a GPU, and announced its weights under Apache 2.0 on Hugging Face. These are provider statements, not Hyperion measurements. Runtime support, memory, security, data flows and operating ownership still need evaluation on the target hardware.

What do we receive from Hyperion?

The engagement names the decision, delivery scope, capacity, acceptance evidence and handover. A Review provides advice and excludes implementation. Any implementation needs an expressly agreed scope, responsibilities and acceptance criteria; ongoing executive leadership requires a written mandate.