CroissantLLM and Lucie: the 100% French models in local

#Why a 100% French model?

Most high-performing open-weight LLMs in 2026 come from China (Qwen, DeepSeek, GLM) or the United States (Llama, Gemma, Phi). They often speak French very well — Qwen 3.8 and Gemma 4 are excellent, incidentally — but they were trained primarily on English and Chinese content, with cultural biases that show through: implicit references, awkwardly translated American phrasing, political perspectives, and an imperfect handling of formal and informal language.

A model trained in France addresses three distinct needs: native quality in formal French, culturally relevant content (law, administration, literature, media), and legal sovereignty in the sense that the weights, datasets, and developers fall under European law. For a law firm, government agency, or public research project, this is often decisive.

#Overview of French LLMs in 2026

Five families dominate today’s self-hostable French landscape. They are not in the same league in terms of size, quality, or licensing.

Mistral (Mistral AI, Paris)

The champion. Mistral Small 24B, Magistral, Devstral. Apache 2.0 for the open models, 128k context, French quality on par with the best cloud models.

CroissantLLM (Centrale Supélec / Illuin)

A truly open-source 1.3B bilingual FR-EN model: weights, datasets, and code. Designed for research and edge deployment.

Lucie (LINAGORA / OpenLLM-France)

7B Apache 2.0, community project, trained on French public supercomputers (Jean Zay). The first true French model of this size.

Pleias (Paris)

Family of 1B to 3B models trained exclusively on open data (Common Corpus). Strict compliance, perfect for the public sector.

Croco / CroissantCool / derivatives

Community fine-tunes for French instructions. Often based on Mistral or CroissantLLM, hosted on Hugging Face.

Alongside them, some non-French models stand out in French: Qwen 3.8 (Alibaba), Gemma 4 (Google), and gpt-oss (OpenAI). We put them into perspective below in the benchmarks.

#CroissantLLM: 1.3B bilingual, fully open

CroissantLLM was created through a collaboration between CentraleSupélec, Illuin Technology, Inria, and INRIA Paris. It is a 1.3-billion-parameter model trained on approximately 3 trillion tokens, split evenly between French and English. Small, certainly, but its distinctive feature is complete transparency: weights, datasets, training recipes, evaluation scripts—everything is published.

Size

1.3B parameters, ~800 MB in Q4. Runs on a CPU, a recent smartphone, or a Raspberry Pi 5.

Languages

Balanced FR-EN bilingual (1:1 in training).

License

MIT, commercial use permitted without restriction.

Variants

CroissantLLMBase (pretrained), CroissantLLMChat (instruction-tuned), CroissantCool (community RLHF).

Usage

Embedded, edge, research, local fine-tuning. Not a general-purpose assistant to pit against GPT-5.

ollama run croissantllm/CroissantLLMChat-v0.1-GGUF:Q4_K_M

>>> Explique-moi la différence entre une SARL et une SAS en 4 lignes.

#Lucie: the French open-source bet

Lucie is LINAGORA's flagship project, developed in partnership with OpenLLM-France and several public research labs. Introduced in early 2025, it is a 7-billion-parameter model trained on the Jean Zay supercomputer at CNRS-IDRIS, on a corpus combining French, English, German, Spanish, Italian, and code, with strong weighting toward French.

Size

7B parameters, ~4.5 GB in Q4_K_M, ~5.5 GB in Q5_K_M.

VRAM

5 GB minimum in Q4. Runs comfortably on RTX 3060 12 GB, or on a 16 GB unified-memory Mac with an M-series chip.

License

Apache 2.0, so it can be used commercially and redistributed.

Context

32k tokens in chat version (Lucie-7B-Instruct).

Sovereignty

Weights hosted on HuggingFace, but the code and data are accessible. Self-hosted mirroring is possible for government agencies.

What stands out in use: Lucie handles French very cleanly, without the “English calque” still noticeable in Llama 4 with literary phrasing. It is, however, a notch below the best models in its tier, such as Qwen 3.5 9B, for code and mathematical reasoning.

ollama pull OpenLLM-France/Lucie-7B-Instruct-gguf:Q4_K_M
ollama run OpenLLM-France/Lucie-7B-Instruct-gguf:Q4_K_M

#Pleias: radical compliance through data

Pleias takes a rare approach: it trains its models only on Common Corpus—a public, traceable dataset with no copyrighted content. The result is 1B and 3B models that an administration can adopt without concern about intellectual-property issues, but whose capabilities remain modest compared with large models trained on the entire web.

Sizes

Pleias-1B and Pleias-3B (Pico, Nano versions). RAG-friendly.

Specialty

Good at RAG, classification, and structured extraction tasks. Weak at free-form chat.

Argument

No training on copyright-protected data. Supply chain auditing possible.

Typical use case

Summarizing public documents, processing administrative requests, and classifying mail.

Pleias is not a conversational assistant; it is an extraction and structuring tool. For a public-sector organization subject to the AI Act and a strict accountability chain, it is probably the only French model that meets all the requirements without compromise.

#Mistral: the French reference

No need to go longer: Mistral remains the French open-weights leader. Mistral Small 24B (Apache 2.0, ollama run mistral-small, 14 GB in Q4) is probably the best self-hostable French model in 2026 on enthusiast hardware. For code, Devstral 24B (Apache 2.0) is the in-house agent specialist. Mistral Magistral brings reasoning (chain of thought) in French, covered in a dedicated guide on the site.

Mistral Small 24B

~14 GB in Q4 (ollama run mistral-small). RTX 4090 or MacBook Pro M3/M4 with 32+ GB. Probably the best French/quality compromise.

Devstral 24B

~14 GB in Q4, Apache 2.0 (ollama run devstral:24b). The code-agent specialist from Mistral AI.

Mistral Magistral 24B

Chain-of-thought reasoning in French.

Codestral 22B

Code-specialized but non-production license: prohibited for work. Prefer Devstral 24B.

#FR quality tests vs. Qwen 3.8 and Gemma 4

Here is a qualitative comparison grid based on typical uses by a French-speaking user (formal writing, PDF summarization, general legal questions, coding). Observations were conducted at temperature 0.3, 8k context, Q4_K_M, with identical prompts.

Mistral Small 24B (FR)

Excellent writing, polished register, rare agreement errors. Very good code. French reference.

gpt-oss 20B (US)

Excellent multilingual performance, with French exceeding expectations. Some English layers in literary text. Very fast (MXFP4).

Qwen 3.8 27B (CN)

Very good formal French, sometimes awkward in casual language. Code at Mistral's level, superior reasoning—consider setting reasoning to low, or it will overthink.

Lucie 7B (FR)

Clean French, with a fluent formal administrative register. Average coding ability. Good for writing and summarization.

Qwen 3.5 9B (CN)

Strong in French for its size (8 GB), 256k context, vision. A good writing and coding compromise on a small configuration.

Gemma 4 12B (US)

Good French, sometimes overly academic. Multimodal, good RAG, Apache 2.0.

CroissantLLM 1.3B (FR)

The level of a lightweight assistant; not in the same category. Excellent for its size.

#Sovereignty and the AI Act

The European AI Act, which came into force in 2024–2026, imposes several obligations on providers and deployers of general-purpose AI models: technical documentation, a summary of training data, labeling of generated content, and systemic-risk assessments for very large models. Models trained and published in the EU simplify auditing, but the AI Act does not say that you need a French model—it says you must be able to document and explain the system.

Local hosting argument

Regardless of the model, running it locally resolves 90% of GDPR concerns. This matters more than the model's nationality.

Transparency argument

For an AI Act audit, a model whose datasets are public (Pleias, CroissantLLM, Lucie) is easier to defend than a model whose training is opaque.

Ecosystem argument

Supporting Mistral, LINAGORA, and Pleias helps consolidate a European industry. It is a political argument, but a real one for public procurement.

Legal argument

The model weights published under Apache 2.0 by a European entity reduce uncertainty around data transfers to the US.

#Which one to choose for your use case

1. 01
   You're just getting started and want to test a French model

   Install Lucie 7B (ollama pull OpenLLM-France/Lucie-7B-Instruct-gguf:Q4_K_M), ~5 GB, a truly 100% French model that runs everywhere. For a stronger general-purpose French model, run ollama run mistral-small (Mistral Small 24B, 14 GB).
2. 02
   You have 12 GB of VRAM (RTX 3060/4070)

   Lucie-7B-Instruct in Q5_K_M for pure French, or Qwen 3.5 9B in Q8 (11 GB) for a more capable generalist. You can also try Mistral Small 24B in Q3.
3. 03
   You have 24 GB (RTX 4090/3090, Mac 32+ GB)

   Mistral Small 24B in Q4_K_M is probably the best self-hostable French LLM, period. Add Lucie 7B for fast “pure French” use cases.
4. 04
   You are a government agency or public-sector organization

   Pleias 1B/3B for document RAG (zero copyright risk), Lucie 7B for general assistance, Mistral Small 24B when you need quality. Avoid models with unclear licensing.
5. 05
   You use embedded / edge / Raspberry Pi

   CroissantLLM 1.3B Q4. It is the smallest French model that remains usable. See the LLM guide for Raspberry Pi for details.

#Go further

Three related paths depending on your next step: to go deeper on French reasoning, the guide dedicated to Mistral Magistral details the FR chain of thought; to build a RAG system over your PDFs with one of these models, the local RAG guide with Ollama without coding covers the complete stack; to frame enterprise compliance, the local LLM and GDPR guide goes through the AI Act obligations point by point.