News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
Lucie-7B

Lucie-7B

Open multilingual language model focused on French and English

Lucie-7B is a 7-billion parameter causal language model built by LINAGORA and the OpenLLM-France community. Released in 2025, it trains on balanced French and English data to reduce anglo-centric bias while maintaining open weights and compliance with OSI definitions.

Share on WhatsApp

  • Trained on 3 trillion tokens with balanced French and English data
  • Uses Llama 3.1 architecture with 6.7 billion parameters
  • Supports context length up to 32,000 tokens after extension
  • One of the first OSI compliant language models
  • Training used 550,000 GPU hours on Jean Zay supercomputer

What it is

Lucie-7B is a pretrained causal language model with approximately 6.7 billion parameters. It was developed by LINAGORA and the OpenLLM-France community to address the lack of high-quality French resources in large language models. The model focuses on balancing French and English data, each comprising roughly one third of the training set. It supports several other European languages and programming code. The project aims to provide an open alternative that respects data rights and reduces cultural bias found in anglo-centric datasets.

How it came about

The model was introduced in March 2025 via a research paper describing both the model and its training dataset. The developers sought to fill gaps in modern datasets by including French cultural heritage documents alongside traditional web sources. The project emphasizes openness, releasing model weights, intermediate checkpoints, and training code on public platforms like Hugging Face and GitHub. This release positions Lucie-7B as one of the first language models compliant with the new Open Source Initiative definition.

Inside the model

The architecture mirrors Llama 3.1, using a decoder-only transformer design. Training occurred in three phases on the Jean Zay supercomputer using 512 H100 GPUs. First, main pre-training processed 3.1 trillion tokens with a 4,096 token context. Next, context extension increased this limit to 32,000 tokens using 5 billion additional tokens. Finally, an annealing phase refined the model on high-quality mixed data. The vocabulary size is 65,024 tokens, and the feed-forward hidden size is 12,288.

What it can do

Lucie-7B handles text generation tasks in multiple languages, with specific strength in French and English due to balanced training data. It also processes several programming languages, which make up nearly 15 percent of the training mix. The model supports instruction following through derived variants like Lucie-7B-Instruct-v1.1. Developers claim it achieves promising results compared to other state-of-the-art models while prioritizing data rights and minimizing copyrighted material in its training set.

How to use it

Model weights are available for download under an OSI compliant licence. The developers provide the full training dataset, processing code, and optimizer checkpoints publicly. Users can load the model using standard libraries, with options for 4-bit quantization to reduce memory requirements. Intermediate checkpoints from various training steps are also accessible, allowing researchers to study model convergence and development over time.

Why it matters

The project demonstrates that open approaches prioritizing data rights can yield competitive performance. By including cultural heritage documents, it offers a resource for preserving and representing French linguistic nuances. It serves as a baseline for future aligned models and encourages the development of less anglo-centric artificial intelligence systems within the European open-source community.

Questions people ask about Lucie-7B

What languages does Lucie-7B support?

The model focuses primarily on French and English, which each make up about 33 percent of the training data. It also supports Spanish, German, Italian, and several programming languages to provide broad multilingual capabilities.

Is Lucie-7B free to use?

Yes, the model weights, training code, and dataset are openly available. It is released under an OSI compliant licence, allowing for broad usage while respecting data rights and minimizing copyrighted material.

What hardware is needed to run Lucie-7B?

For efficient inference using 4-bit quantization, you need at least 6GB of VRAM on a GPU. Without quantization, the requirement increases to 16GB of VRAM. The model was trained on H100 GPUs but can run on consumer hardware with quantization.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Lucie-7B yet, so the facts on this page come from the sources listed above: huggingface.co/OpenLLM-France/Lucie-7B, arxiv.org/abs/2503.12294. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.