Built from scratch. Built in Europe.

Olala Mini 7A1B is the new range of foundation LLMs from OVHai LLM: sovereign, frugal, open and responsible.

Our story: European and built from scratch

A European model, from the first line of code

In November 2026, OVHai LLM launched Olala, a new range of models designed and built from the ground up in France. The result of years of experience and made possible with the help of Europe, Olala puts sovereignty at its core. The first model, Olala mini, was trained on three public EuroHPC supercomputers: Jupiter, MeluXina and Jean Zay.

The range consists of 5 models:

Olala v1.0 Translate

A translation model based on Olala mini 7A1B able to translate your content in 18 languages.

Olala v1.0 mini 7A1B

The generalist model, finetuned and trained with reinforcement learning.

Olala v1.0 Summarise

A summarisation model based on Olala mini 7A1B able to produce a structured summary of your textual contents.

Olala v1.0 Base

The base model of Olala mini.

Olala v1.0 SFT

The supervised fine-tuned checkpoint from which Olala-v1.0-mini-7A1B was trained with reinforcement learning.

Performance: small and powerful

Small footprint. Serious performance.

Olala Mini delivers competitive results at a fraction of the size and energy cost of larger models. The model relies on a 7 billion parameters architecture but only a fraction of the model is used for each request, making it faster and less memory-hungry.
This allows Olala mini to run on GPUs with as low as 8GB of RAM such as an RTX A4500, or even without any GPU, on a standard CPU. Olala mini can thus run where you need it:

On-premise

The model can run on your servers, with or without GPUs.

On a laptop

The model can run on your laptop, with or without GPUs.

At the edge

The model can be trained for very specific tasks to run on very small devices.

Benchmarks

Olala-v1.0-mini-7A1B: benchmark scores

Scores in %. Bold = best on the benchmark, underlined = second best. Same 19 models and order as the model card, grouped by size.

Benchmark Olala-v1.0-mini-7A1B Granite-4.0-H-Tiny LFM2.5-8B-A1B Ling-3.0-tiny OLMoE-1B-7B-Instruct LFM3-1.7B Qwen3.5-2B LFM2.5-2.6B Ministral-3-3B-Reasoning Qwen3.5-4B SmolLM3-3B Apertus-8B-Instruct EuroLLM-9B-Instruct Falcon-H1R-7B Ministral-3-8B-Reasoning Mistral-7B-Instruct-v0.3 Olmo-3-7B-Think Luciole-23B-Instruct Luciole-23B-Thinking
Size group 7-8B MoE, ~1B active Up to 2B dense 2.6-4B dense 7-9B dense 23B dense
Params 7B-A1B 7B-A1B 8B-A1B 7.9B-A1.3B 7B-A1B 1.7B 2B 2.6B 3B 4B 3B 8B 9B 7B 8B 7B 7B 23B 23B
Instruction following
IFBench 57.7 14.3 48.3 57.3 17.0 70.0 35.3 59.0 33.7 49.7 19.0 19.3 15.0 45.3 27.0 18.0 38.7 26.3 23.7
Multi-IF 71.2 44.5 71.0 75.2 25.0 83.6 57.3 77.7 48.6 71.8 42.9 41.8 37.7 65.7 49.7 25.5 61.2 51.6 54.6
LIFEBench 79.3 63.4 77.8 82.3 51.0 82.7 82.0 79.3 68.4 84.0 64.6 62.4 64.0 72.1 72.8 58.2 71.1 61.5 64.3
Knowledge and reading
GPQA Diamond 50.5 30.3 41.9 68.7 15.7 62.6 63.1 56.1 42.4 75.8 42.9 23.7 30.3 62.1 56.6 21.2 48.5 28.8 36.4
TruthfulQA 66.6 48.7 68.9 66.5 12.7 65.6 60.5 62.9 47.7 74.7 59.9 21.7 47.4 67.3 63.4 8.0 65.6 57.6 64.3
DROP 83.8 64.7 73.1 87.6 36.4 85.5 81.4 79.6 82.2 86.3 54.0 55.0 62.5 85.7 85.6 35.8 86.7 63.4 78.7
Code
HumanEval+ 88.3 77.4 79.9 92.0 32.4 92.2 48.3 90.2 92.6 83.4 81.3 34.6 36.0 93.4 93.8 32.8 91.1 44.1 77.4
MBPP+ 75.1 72.8 64.2 79.1 26.0 80.8 49.0 74.9 78.5 73.5 70.8 44.3 48.6 82.0 81.0 45.7 79.0 62.6 71.7
Math
AIME 2024 60.7 4.0 64.7 78.7 0.7 81.3 24.0 66.7 74.7 80.7 46.0 0.0 0.0 94.0 78.0 0.0 69.3 10.0 30.0
AIME 2025 61.3 4.7 48.0 70.0 0.0 79.3 26.7 51.3 70.0 74.7 32.7 0.0 0.7 82.7 78.0 0.0 64.0 13.3 31.3
AIME 2026 52.7 10.0 48.0 75.3 0.0 80.0 27.3 57.3 73.3 82.0 45.3 0.7 0.7 90.0 82.7 0.0 67.3 6.0 23.3
MGSM 68.7 62.0 60.5 73.7 22.0 79.3 73.3 75.7 77.7 88.8 61.7 40.7 39.8 66.9 84.8 28.7 71.7 45.9 61.0
Tool use
BFCL V3 59.6 60.2 69.7 68.3 n/a 74.1 59.2 67.9 37.7 67.8 42.6 n/a n/a 67.8 42.0 48.2 n/a 60.4 35.1
Luciole-23B-Thinking is an interim checkpoint that is not publicly released. OLMoE-1B-7B has a 4,096-token context window; prompts that do not fit count as wrong. n/a: the model has no native tool-call support.
All benchmarks are entirely replicable. The information is available in the technical report.

Olala-v1.0-summarise: summarisation benchmark

ROUGE-Lsum F1 x 100 (stemmed) on 11 public test sets: up to 1,000 documents each (seed 42), documents cut at 12K words, one-turn prompt <instruction>\n\n<document>. Average = plain average over the 11 test sets. Bold = best, underlined = second best. Length ratio = summary words / reference words.

Test set Olala-v1.0-summarise OLMoE-1B-7B-0924-Instruct† LFM2-8B-A1B LFM2.5-8B-A1B‡ LFM2.5-1.2B-Instruct LFM3-1.7B Gemma-4-E2B-it Qwen3.5-2B LFM2-2.6B Falcon-H1-3B-Instruct SmolLM3-3B NVIDIA-Nemotron-3-Nano-4B Qwen3-4B-Instruct-2507 Gemma-3-4B-it Falcon-H1-7B-Instruct Mistral-7B-Instruct-v0.3 Granite-4.1-8B
Size group 7-8B MoE, ~1B active Up to 2B dense 2.6-4B dense 7-9B dense
Params 7B-A1B 7B-A1B 8B-A1B 8B-A1B 1.2B 1.7B 2B effective 2B 2.6B 3B 3B 4B 4B 4B 7B 7B 8B
News
CNN/DailyMail 33.0 29.1 28.9 28.2 32.9 28.0 30.6 29.8 29.0 34.1 32.5 31.7 29.5 27.1 33.5 31.8 29.9
XSum 16.5 16.9 16.4 15.7 16.9 14.8 17.5 16.3 16.9 22.4 16.7 16.7 18.0 18.6 21.8 18.3 18.4
Multi-News 36.3 36.4 33.7 32.5 33.9 34.6 34.1 37.0 37.0 37.6 39.1 36.6 36.9 36.1 37.2 38.6 37.6
Dialogue
SAMSum 33.4 30.6 28.9 29.6 29.4 30.1 32.2 29.6 30.1 31.5 33.1 33.1 32.9 28.9 32.8 32.5 33.1
DialogSum 28.1 21.3 24.2 24.3 23.2 21.5 23.3 24.7 24.3 24.4 25.0 27.1 26.4 21.6 26.5 24.7 27.6
Forum posts
TL;DR 21.2 21.7 16.3 15.7 18.2 15.7 19.0 15.6 16.2 18.5 17.9 18.1 18.2 17.5 25.2 18.7 17.9
Earnings calls
ECTSum 15.9 9.4 12.1 12.5 13.4 11.3 13.2 13.6 13.1 15.0 13.2 13.7 11.5 13.1 15.4 14.5 12.4
Government and legal
BillSum 38.8 38.6 41.0 35.7 43.6 32.9 37.2 35.0 41.0 40.0 38.8 37.3 31.7 35.2 40.3 44.5 37.0
GovReport 43.0 45.8 42.7 41.4 44.2 44.3 49.6 49.9 46.5 48.0 51.4 47.4 49.1 45.7 49.4 49.6 51.1
Science
arXiv 37.4 33.6 36.0 34.7 35.4 33.1 35.6 34.2 36.2 38.9 25.9 37.0 35.1 35.5 37.8 37.5 32.8
PubMed 40.5 37.4 38.0 36.1 37.6 36.1 41.1 40.5 39.7 43.6 31.2 40.0 40.5 38.4 43.5 42.6 40.4
Average 31.27 29.17 28.94 27.86 29.87 27.47 30.31 29.66 29.99 32.17 29.52 30.78 29.98 28.88 33.04 32.14 30.76
Length ratio 2.32 4.35 2.74 2.67 1.89 3.67 3.31 3.43 2.90 2.69 4.41 2.63 3.70 3.23 2.69 3.12 3.78
Olala-v1.0-summarise: reasoning_effort="medium", temperature 0.8, top_p 0.9, top_k 40. Other models: no thinking, greedy decoding, max 4,096 new tokens.
† 4K-context model: long documents are cut to fit. ‡ Reasons even with thinking off; the reasoning is removed before scoring.
All benchmarks are entirely replicable. The information is available in the technical report.

Translation leaderboard

10 models evaluated on two translation benchmarks, scored with BLEU. Click a column header to sort.

28 directions, BLEU Financial documents, BLEU
Model Size WMT newstest DOLFIN en→fr
Olala-v1.0-translate 7B-A1B 32.6 62.6
Gemma 4 E4B 4B 31.9 54.0
Gemma 3 4B 4B 28.7 51.9
Qwen3.5 4B 4B 29.6 51.7
Hy-MT2 7B 7B 29.1 51.7
Tower+ 2B 2B 32.9 50.8
TranslateGemma 4B 4B 26.8 48.6
Hy-MT2 1.8B 1.8B 25.2 47.5
Qwen3 4B 2507 4B 26.9 46.8
Ministral 3 3B 3B 20.9 36.7
Scoring. Scores are BLEU; the best score in each column is in bold.
All benchmarks are entirely replicable. The information is available in the technical report.

Other metrics

How much CO₂ does Olala mini emit per million tokens?

One million tokens is roughly 750,000 English words. Here is what generating them with Olala 1.0 mini emits, by machine and load.

Machine Load g CO₂e per million tokens
Server with an H100 GPU 32 or more requests at once 12est.
Laptop (MacBook M1 Pro) One request at a time 30
Server with an H100 GPU 4 requests at once 42
PC with 8 GB of RAM and no GPU (AMD Ryzen 7700) One request at a time 65est.
Server with an L4 GPU 4 requests at once 101
Server with an H100 GPU One request at a time 140est.
Electricity and hardware manufacturing, whole machine, in France. est. = estimated; other values are measured.

At most, one million tokens emits about as much as driving a petrol car for one kilometre. Load matters more than hardware: a busy H100 server emits about ten times less per token than one serving a single request, and an idle reserved server still emits almost as much per hour as a busy one. For a single user, a laptop or a PC without a GPU is the low-carbon option. In the United States, the same server would emit about three times more.

Method Boavizta method (API 2.6): electricity × French grid intensity (0.098 kg CO₂e/kWh) × 1.2 for data-centre cooling, plus the share of hardware manufacturing for the time used. Server figures include the whole machine, not only the GPU. Measurements: OVHcloud, Gravelines, November 2026, GPU power read during the runs. Speeds are measured on the laptop (65 tokens/s) and the PC (48 tokens/s); their power draw is assumed (22 W and about 65 W). Only generated tokens are counted.

What has been done to measure and reduce harmful or toxic contents?

Between SFT and the final model, we ran reinforcement learning (RL) in four stages: instruction following and math, code, assistant behavior, and finally agentic tasks. For each prompt, the model generates several responses, and the ones with the highest score are reinforced. The score comes from a program when the answer can be checked automatically, and from an LLM judge otherwise.

Safety is present to some extent from the first stage, but most of the work happens in the “assistant behavior” stage, using two datasets:

  • 23,754 prompts drawn from public jailbreak and prompt-injection benchmarks, plus scenarios where the user tries to get around a system prompt’s instructions, sometimes over multiple turns (repeated asks, gradual escalation). They span five languages and mix attacks with legitimate requests;
  • 1,112 cases, each scored against its own rubric. These notably include the weaknesses found through manual red-teaming: role-play, professional pretexts, and multi-turn escalation.

The key point is how responses are scored. The judge penalizes both helping with harmful use and over-refusal. And nearly half of the prompts are requests where the model should help. So the model can’t “win” by refusing everything.

Metric Olala-v1.0-mini-7A1B-SFT Olala-v1.0-mini-7A1B
Attacks n = 5,701
Guard ASR (%)↓ 12.5 6.3
Guard ASR, strict (%)↓ 24.6 13.5
Harmful, rubric judge (%)↓ 34.9 22.2
Benign n = 299
Over-refusal (%)↓ 7.0 10.7
All prompts
Reasoning overflow (%)↓ 11.8 15.0
↓ lower is better. Bold = better of the two.

That explains the shape of the results: far fewer harmful responses, including outside English (likely thanks to the multilingual data), with only a modest increase in over-refusals. The other stages (code, agentic) may also have played a role, since the comparison covers the full RL cascade.

What can I do with Olala mini?

Two examples of tasks for Olala mini

Most of what an AI agent does all day is not grand reasoning. It is a steady flow of small, well-defined steps, repeated thousands of times. That is where a small, fast, frugal model earns its place.

Case 1

One thing at a time

  • Tool calling
  • SQL
  • Data extraction
  • Structured output
  • Classification
  • EN / FR

Look inside a production agent and you find the same few moves again and again: query a database, read the result, pull a date or an amount out of an invoice, fill in a JSON form, run a calculation, check a rule, sort a request into the right category, call the right connector, answer from a document with an exact quote. We built a small benchmark around these everyday moves, in English and French, and scored it automatically against known answers, with no AI model acting as judge. One rule applies throughout: Olala mini never answers about data from memory. It reads through tools (SQL, storage) and computes through a calculator, so every answer can be traced back to a tool result. That is how reliable agents work, and each of these calls is cheap enough to run thousands of times a day.

Case 2

Choosing the right beam

  • Multi-step workflow
  • Python sandbox
  • Catalogue lookup
  • Engineering calculation
  • Verification

Our first teaser showed a strength-of-materials calculation, so we took it literally. We gave Olala mini a project note, a supplier catalogue and a Python sandbox, and asked it to find the cheapest beam that carries the load without bending too much. An engineer does this in minutes: read the requirements, filter the catalogue by location, then check each candidate from cheapest up. Most agent work has this same shape: retrieve, extract, query, compute, verify, repeat. Olala mini handles the individual steps well. It finds the right note, extracts the load, span and location, and queries the catalogue correctly. Inside a workflow that splits the job, hands off each step and checks every result, it completes the shorter files from start to finish. This is the setup we recommend: the workflow runs the process, Olala mini does the steps, and code checks the numbers.

Transparency: released open weight

Open by design

We release Olala openly so you can inspect it, test it and build on it. No black box.
Olala mini is released with its weights, its pre-trained base model, its technical report and, of course, the full AI Act compliance documentation.