Lucie 7B

Lucie-7B is the first open foundation model from the OpenLLM-France consortium, led by Linagora with support from CNRS and GENCI's Jean Zay supercomputer under the France 2030 program. The model and its training dataset were built to offset the anglocentric bias of most pretraining corpora: alongside standard web data, Lucie was trained on French cultural heritage documents, with French as the central language of a multilingual mix.

The honest picture: Lucie's January 2025 chatbot launch was a public fiasco — the raw, non-RLHF-tuned model gave absurd answers (cow's eggs, botched arithmetic) and was pulled offline within three days. Treat Lucie-7B as what it is: an academic research artifact and a milestone for European AI sovereignty, not a production-ready assistant. The open dataset is arguably the more valuable contribution.