Fine-tuning service

Olala mini fine-tuning

Your business task, your specialised model

We scope your use case, build the benchmark, prepare your data, train, evaluate, deploy and keep the model current. One team, one order, from first meeting to production.

Launch: Olala mini at the OVHcloud AI Summit · 19 November 2026

Why fine-tuning needs experts around the GPU

Fine-tuning is the easy part. The expertise is everything around the GPU.

Training a 7B model takes a few hours of GPU time. What takes weeks, and decides whether your project succeeds, is the work around it:

  • Choosing the right approach: deciding whether fine-tuning is the best answer to your problem, or whether another option would serve you better.
  • Building a reliable benchmark: defining how success will be measured before any training starts.
  • Preparing your data: turning scattered, unstructured documents into clean, high-quality training examples.
  • Proving it works: validating the results on your real use cases, not just on test sets.

OVHcloud provides the training and serving tools. Our experts run them and handle everything around them. We are the team that designed Olala mini, so we bring that experience to your project.

Six stages, all run by OVHai LLM experts

The fine-tuning pipeline

Our team carries out every stage. For training and serving, we operate OVHcloud AI Training and OVHcloud AI Deploy. You place one order, with OVHai LLM.

OVHai LLM experts
OVHcloud tools
01

Scope

Use case, success criteria and the benchmark, built from your real cases.

02

Prepare data

Collection, cleaning, anonymisation, expert-validated and synthetic examples.

03

Train

Fine-tuning runs with recipes tuned for Olala mini.

OVHcloud AI Training

GPU training jobs, operated by our team

04

Evaluate

Scores on the benchmark, frontier comparison, non-regression and safety.

05

Deploy

Private endpoint in Europe, or on your infrastructure.

OVHcloud AI Deploy

Managed endpoints with autoscaling, operated by our team

06

Maintain

Re-evaluation, drift alerts, retraining, new Olala versions.

Stages 02 to 04 loop: each evaluation on the benchmark shows what to fix in the data before the next run.

Built during scoping, before any training, and used at every stage

The benchmark comes first.

We build the benchmark with your business experts from the scoping stage. It turns “good enough” into measurable criteria, shows whether fine-tuning is worth it, and makes every training run comparable.

  1. Scoping

    Define success criteria with your experts

  2. Build

    100 to 200 real cases with expected answers

  3. Baseline

    Olala mini as is, tuned prompts, frontier model, RAG

  4. Training

    Every run scored on the same benchmark

  5. Production

    Final report, then periodic re-evaluation

Sprint 0 covers scoping, benchmark and baseline.

Where every project starts

Sprint 0: measure before you train

No one can tell you how much fine-tuning will improve your results until they’ve seen your data. Sprint 0 replaces guesswork with measurements.

What we do

  • Frame the use case and agree measurable success criteria
  • Build the benchmark from 100 to 200 of your real cases
  • Measure a baseline: Olala mini as is, with tuned prompts, a frontier model, RAG where relevant
  • Classify errors: format and vocabulary, missing knowledge, reasoning
  • Audit available data: volume, quality, confidentiality

Priced on the days it takes, which depend on the number of use cases and how accessible your data is.

Sprint 0 · Diagnostic

From €5,000 excl. VAT

Final price set on the scope agreed at our first meeting

What you receive

  • Diagnostic report with the measured baseline
  • The benchmark built for your use case
  • A recommendation and a scoped proposal for the next stages

Three possible verdicts

GoFine-tuning can close a measured gap
Another routePrompting or RAG does the job better
Not yetData must be built up first

After Sprint 0

Each next stage is scoped and quoted from what Sprint 0 measured

Option

Proof of value

A first training run on a few hundred examples. Real scores on the benchmark and a go / no-go before the full project.

Project

Full fine-tuning

Complete data preparation, training on OVHcloud AI Training and deployment on OVHcloud AI Deploy, run by our team, with a final evaluation report.

Subscription

Maintenance

Scheduled re-evaluation on the benchmark, drift alerts, retraining on new data, migration to new Olala versions.

One point of contact: OVHai LLM takes the order and stays accountable from scoping to maintenance.

OVHai LLM, the AI models and AI systems unit of the OVH group

The team that designed Olala mini.

We know how Olala mini was trained, which recipes work for fine-tuning it, and where the pitfalls are.

15+years building language models
1,000+models trained and in production
40+corporate clients served
2024winner of the European Commission’s Large AI Grand Challenge

Data in Europe

Processed on OVHcloud infrastructure in Europe

No reuse of your data

Never used to train other clients’ models

One team

Order, project and support with OVHai LLM

Three ways to get started

Let’s build together

1

Bring us a use case

A business task where a specialised model makes the difference

2

Run a Sprint 0

Benchmark, baseline and a clear verdict, from €5,000 excl. VAT

3

Meet us at the AI Summit

19 November 2026, launch of Olala mini

Contact

Gilles Montier

contact@llm.ovhai.com