Ranking updated on 09/10/2026
Ranking of open-weight LLMs capable of producing polished, idiomatic French without anglicisms. We favor models with dense French-language corpora (French or European origin) and a free license to facilitate enterprise deployment.
First open Mistral reasoner. AIME24 70.7%. Based on Small 3.1 + CoT training.
ollama run magistral:24b
June 2025 update for Small 3.1. Half as many infinite generations, improved function calling.
ollama run mistral-small3.2:24b
Small 3 enhanced with vision. 128k ctx, Apache 2.0. Replaced by Small 3.2 since June 2025.
ollama run mistral-small3.1:24b
Sovereign French-language LLM, trained on French corpora.
ollama run lucie:7b
Co-developed with NVIDIA. 128k ctx, Tekken tokenizer, strong in European multilingual.
ollama run mistral-nemo:12b
Excellent quality-to-size ratio at release (early 2025). Competes with 70B models.
ollama run mistral-small:24b
LoRA-finetuned Falcon-7B on spontaneous French dialogue. ⚠ NC-SA license. Separate Apache variant.
# HuggingFace : OpenLLM-France/Claire-7B-0.1 (Apache : Claire-7B-Apache-0.1)
French. Trained 100% on Common Corpus (open data). EU AI Act compliant by design.
# HuggingFace : PleIAs/Pleias-3b-Preview
| Rank | Model | Params | Q4 VRAM | Context | License |
|---|---|---|---|---|---|
| #1 | Magistral Small 24B | 24B | 14 GB | 128 000 | Apache 2.0 |
| #2 | Mistral Small 3.2 24B | 24B | 14 GB | 128 000 | Apache 2.0 |
| #3 | Mistral Small 3.1 24B | 24B | 14 GB | 128 000 | Apache 2.0 |
| #4 | Lucie 7B | 7B | 5 GB | 4 096 | Apache 2.0 |
| #5 | Mistral Nemo 12B Instruct | 12B | 7 GB | 128 000 | Apache 2.0 |
| #6 | Mistral Small 3 | 24B | 14 GB | 32 768 | Apache 2.0 |
| #7 | Claire 7B 0.1 | 7B | 5 GB | 2 048 | CC-BY-NC-SA 4.0 |
| #8 | Pleias 3B Preview | 3B | 2 GB | 2 048 | Apache 2.0 |
Two options depending on the memory you need and your software: Radeon RX 9070 XT 16 GB · GMKtec EVO-X2 64 GB / 1 TB (Ryzen AI Max+ 395) — compare AI hardware →
You found the best French model. The Local AI Kit explains when to prefer it over a general-purpose model (ch. 10) and gives you writing, translation, and summarization instructions that actually work (ch. 9).
We keep the models of French/European origin (Mistral, Lucie, CroissantLLM, Teuken, EuroLLM) and the large general-purpose models with very good French (Qwen, Llama). Bonus: European origin (sovereignty), size ≥ 24B (literary ability), Apache/MIT license.
Criteria considered:
The scoring is fully transparent: see our methodology for details on VRAM/tokens/sec calculations.
What is the best LLM in French in 2026?
Magistral Small 24B is our #1—it benefits from a dense French-language training corpus and a France-based team. But depending on your hardware, a lighter model (Lucie 7B, Mistral 7B) may be preferable.
Lucie vs Mistral 7B: which one should you choose?
Lucie 7B is 100% sovereign (transparent corpus, trained at Jean Zay/CNRS), but its context is limited to 4k tokens and it is less versatile in English. See the full comparison.
Is there a high-performing non-French European LLM?
Yes — Teuken 7B Instruct (🇩🇪 OpenGPT-X) covers the EU's 24 languages and has a commercial Apache 2.0 license. EuroLLM 9B (EU consortium) is also solid. Comparison.
Which LLM for legal documents in French?
For French legal work, language quality comes first. Mistral Small 3.1 24B or Mistral Large 3 are the best choices if your hardware can handle them. Pair it with a Local RAG to inject your legal texts.
Learn more with our detailed head-to-head matchups of the finalists: