Free artificial intelligence for the languages of France
Project in developmentLengo nostro · Our language
Giving the regional languages of France a free and sovereign digital voice.
Lingua Nostra, from Latin for "our language"; lengo nostro is one of the ways Provençals name their own.
The project
Lingua Nostra is a project to train free, self-hostable generative AI language models (LLMs) for the regional languages of France — to understand, write, and speak them. Lou Pichoun, the first open-source Provençal language model, will be the first to see the light of day, born of Provence and carried by it. Breton, Corsican, and Alsatian may follow.
The ambition: to give each language a digital presence it does not yet have.
Provençal is not an end in itself: it is the pilot project. Lou Pichoun will serve as a methodological laboratory, and its lessons will be adaptable to the other regional languages in turn.
Read the full projectWhere things stand
Lingua Nostra is, at this stage, a research project at the concept level. No prototype exists yet — and that is the logical consequence of where we are: our immediate priority is building the corpus and finding the first scientific and financial partners.
Before developing a reliable AI for a regional language, one must gather a quality corpus, obtain the necessary permissions, and assemble the scientific and technical resources. The absence of a demonstration is therefore not a shortcoming: it marks the starting point of rigorous work.
See the roadmapWhy
Irish
Basque
Catalan
A handful of minority languages have already entered artificial intelligence: Irish, with a model developed at University College Cork; Basque, with Latxa; and even Catalan. Provençal has no model at all — it would be a first. And most of France's regional languages have nothing.
Lingua Nostra would hand the keys of the digital world to those who want to write, create, publish, and transmit in their own language, rebalancing the place of these languages in the digital space.
A global movement
Across the world, low-resource languages are entering artificial intelligence. Basque already has its model, Latxa, and Irish has its own; even a major official language like Catalan has one, and Northern Sami achieved it with a tiny corpus. In France, the COLaF programme, led by Inria, is building corpora and tools for the country's languages, including Alsatian and Breton.
Provençal still has nothing. Lingua Nostra thus addresses an international scientific challenge — linguistic diversity in the digital world — and would fill a gap among the languages of France.
Why Provençal
Feasibility
Adapting a large AI model to a low-resource language is not a hypothesis. Several teams have already done it, sometimes with very little data and modest hardware.
Adapted from roughly 274 million words on near-consumer hardware. 89% accuracy in comprehension.
Adaptation to a Celtic language structurally distant from English, validated by native speakers.
Adapted with only 22 million words. Usable results despite an extremely small corpus.
| Project | Language | Volume | Result |
|---|---|---|---|
| Northern Sami | Sami | ~22 million words | usable despite a tiny corpus |
| Qomhra | Irish | ~44 hours of compute (2 GPUs) | validated by native speakers |
| LlamaTurk | Turkish | ~274 million words | 89% accuracy in comprehension |
| Lou Pichoun | Provençal | 50 to 100 million words targeted | to come |
The challenge
A language absent from today's artificial intelligence will fade from tomorrow's digital space.
Sovereignty is not only about voice: it extends to writing, translation, and the language as a whole. Lingua Nostra will build models that the community retains control over — informed consent from speakers, open and self-hosted models, no dependence on any proprietary solution.
In practice, "free" means publishing the model and code under a permissive licence, opening corpus data wherever rights allow, and placing governance in the hands of the language community. The tool belongs to those who speak the language, not to any institution.
It is also a matter of digital independence: control over speakers' data, the longevity of a tool no one can take away, and the autonomy of linguistic communities in the face of closed solutions. A language that depends on no one else to exist in the digital world.
In practice, sovereignty changes everything:
« Se tèn sa lengo, tèn la clau
que di cadeno lou deliéuro. »
"Who holds his language holds the key that frees him from his chains."
Frédéric Mistral, I Troubaire catalan, 1861
Lou Pichoun, Lingua Nostra's Provençal language model
Lou Pichoun ("the little one") will be the first open-source Provençal language model (LLM): a generative AI designed to produce, translate, and converse in the Provençal of Provence. Sovereign, it belongs to the community, not to any institution. The long-term ambition is for it to read aloud as well, through a free Provençal text-to-speech synthesis. The project is in its launch phase.
Discover Lou Pichoun in detail →What it would make possible
A free Provençal model is not an end in itself: it is a starting point for those who want to live the language day to day.
Enabling Provençal institutions to write and communicate in the regional language.
Helping artists create and share work in Provençal more easily, thereby keeping the language alive.
Writing correctly in Provençal in everyday life: naming a boat or a home, drafting a note or a sign.
Supporting bilingual schools and Provençal teachers: exercises, correction, conversation.
Equipping institutions and associations working to develop the Provençal language.
Producing more Provençal content online, at scale, while maintaining a high level of quality.
Long-term vision
Lou Pichoun is only a beginning. If the method delivers on its promise, what Lingua Nostra could become in five or ten years goes far beyond a single model:
Who leads the project
Provençal by origin, based in Montréal (French citizen, Canadian permanent resident). Master's degree in Sociology from Université Paul-Valéry Montpellier III, with a thesis on collective Provençal memory.
Consultant in AI strategy and implementation: agent design, content pipelines, language model integration.
At the crossroads of these two worlds: a background in sociology that makes me see languages as living heritage, and experience in artificial intelligence that lets me design and adapt their models. Provençal by origin, I carry this language as much as I research it.
Academic framework
Research is at the heart of Lingua Nostra. The project is intended to be carried out within a doctoral framework, as an independent candidate with academic support, once funding is in place.
The scientific question is interdisciplinary: how can the creation of a free, self-hosted LLM contribute to the digital vitality of a language of southern France — Provençal — and what method of adapting large models makes it achievable for a language with strong dialectal variation?
Three original contributions follow from this: the first adaptation of a model to a language of southern France, managing two dialects in parallel — Rhodanian and Maritime; a study of the transferability of a method already proven on other low-resource languages; and the opening of a methodological path that could eventually be transposed to other languages of France.
Get involved
Lingua Nostra speaks to different audiences, each with a role to play: researchers and universities, local authorities and cultural institutions, patrons and foundations, associations and language communities, technical partners and engineers.
Waiting list
Lou Pichoun is being built. Leave your address to be notified of the model's launch and key milestones.
Would you like to support the project, become a partner, or contribute scientifically? Everything is detailed on the full project page.
Do you share these values, or can you contribute to or support the project? Write to us.
contact@linguanostra.orgFrequently asked questions
Does a chatbot or AI that speaks Provençal already exist?
That is exactly what Lou Pichoun is — the AI chatbot and language model from Lingua Nostra for Provençal. The project is currently being built: sign up for the waiting list to follow its progress.
What is Lou Pichoun?
An open-source Provençal language model (LLM) — a generative AI dedicated to the Provençal of Provence. Lou Pichoun will power a chatbot capable of understanding and producing written Provençal. It is also the first building block of a sovereign AI for natural language processing (NLP) in Provençal.
Will Lou Pichoun be able to translate or conjugate in Provençal?
Yes, and much more. Beyond translating from French into Provençal and conjugating verbs, Lou Pichoun will understand your questions and answer them, exchange ideas and converse, help with drafting and proofreading, and explain a word or a turn of phrase — a genuine conversational assistant, but in Provençal. In the long run, Lou Pichoun will not only write in Provençal: it will read aloud through a free and sovereign Provençal text-to-speech synthesis. The model is in development; these features will arrive progressively.
How can I be notified of the launch?
Sign up for the waiting list: you will receive project announcements.
Join the waiting list
How
The envisaged method, in three steps
Digitising Provençal texts (journals, dictionaries, literature), then cleaning and deduplicating them. Target: between 50 and 100 million words.
Starting from a large multilingual model and adapting it to Provençal in three stages: continued pre-training on the corpus, fine-tuning with instructions corrected by native speakers, then alignment with their preferences.
A network of native speakers validates and corrects the model's responses, choosing between two versions to refine it, via a simple interface. Without them, there is no quality model: they are the sole source of truth.