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

Luciole

Linagora
open source / Overall score: 1.7

Luciole is Linagora and OpenLLM-France's fully open French-English model family and the successor to Lucie, released at 1B, 8B and 23B in base and instruct forms. The weights ship with a 4.65-trillion-token training dataset drawn from openly licensed corpora, the training pipeline and training checkpoints. Base models are pretrained on about 5 trillion tokens, roughly 42 percent English and 30 percent French alongside six other languages.

Openness

5 high confidence
5.0
weights
open(Apache-2.0, safetensors on HF for 1B/8B/23B base + Instruct-1.1)
data
open(Luciole-Training-Dataset on HF, ~4.65T tokens, "only corpora under open licenses")
code
open(Luciole-Training pretraining + data-processing pipeline on GitHub, GPL-3.0)
checkpoints
open(intermediate + final pretraining checkpoints released)
license
Apache-2.0(OSI)

Luciole ships Apache-2.0 weights with its full training stack, including the CC-BY-SA-4.0 Luciole-Training-Dataset of openly licensed corpora, a GPL-3.0 training pipeline, and checkpoints from every training phase that the model card links on Linagora's download server. Unlike Lucie, whose whole corpus was non-commercial, the corpus license permits commercial reuse, though a few bundled sources such as the Claire dialogue sets keep their own non-commercial terms.

Adoption

1 medium confidence
1.0

Adoption is measured as Hugging Face downloads summed across six checkpoints spanning the 1B, 8B and 23B sizes in base and instruct form. That total reflects a niche audience for a sovereign French-language model rather than broad usage.

Capability

2 low confidence
2.0

Luciole is a multilingual French-English family up to 23B parameters that favors French cultural representation and open data rights over benchmark performance, alongside Lucie-7B and Alia-40b. Its 23B-Base card publishes training-convergence and evaluation curves but no comparison against outside models, so this reading comes from scale and positioning rather than a measured benchmark.

Verified 2026-09-27, except some scores, still awaiting re-confirmation