AI Potluck
Contribute
Back to Gap Map Model components / Base / pretrained models

Lucie 7B

Linagora
open source / Overall score: 1.7

Linagora and OpenLLM-France's fully open multilingual 7B foundation model on the Llama 3.1 architecture, trained on roughly three trillion tokens with about equal French and English shares plus German, Spanish, Italian and code. The weights ship with the Lucie Training Dataset, the Megatron-DeepSpeed training code and intermediate checkpoints as repository branches.

Instruct variants are folded into this record; the successor Luciole family is a separate entry.

Openness

5 high confidence
5.0
weights
open(Apache-2.0)
data
open(Lucie-Training-Dataset on HF, redistributed in training form)
code
open(Lucie-Training Megatron-DeepSpeed fork)
checkpoints
open(intermediate HF revisions)
license
Apache-2.0(OSI)

Lucie-7B publishes Apache-2.0 weights with its full training stack: the Lucie-Training-Dataset corpus, the Megatron-DeepSpeed training code, and intermediate checkpoints as repository branches, matching OLMo and Apertus's openness tier. The corpus itself is CC-BY-NC-SA-4.0, published and ungated but not commercially reusable, short of the authors' own OSI-compliance claim.

Adoption

1 medium confidence
1.0

Adoption is measured as Hugging Face downloads summed across the base and instruct checkpoints.

Capability

2 medium confidence
2.0

Lucie-7B is a 7B multilingual base prioritizing French cultural representation and data-rights constraints, competitive for its size and language mix on French/English evaluations in the paper. It remains well below the 2026 open-weight frontier of trillion-parameter mixture-of-experts models.

  • https://arxiv.org/abs/2503.12294 recorded 2026-09-29

    Abstract: the instruct models "achieve promising results compared to state-of-the-art models", presented as "an initial step toward developing more performant, aligned models in the near future".

  • https://huggingface.co/OpenLLM-France/Lucie-7B recorded 2026-09-29

    Model card evaluation section: "Multiple evaluations were conducted during Lucie-7B's training to assess its performance on standard benchmarks", with French, English and multilingual ARC learning-curve figures.

  • https://huggingface.co/OpenLLM-France/Lucie-7B/raw/main/README.md recorded 2026-09-29

    Raw model card: "Main results are summarized in the following figures" for French, English and multilingual ARC, plus needle-in-a-haystack plots; "Lucie-7B has the same neural network architecture as" Llama 3.1 with 6.7 billion parameters.

Verified 2026-09-27