Atomic molecular dynamics (MD) simulations are essential for understanding the dynamics of biological systems. However, achieving accurate simulations remains a challenge. This work presents FeNNix-Biol, a new foundational model based on neural networks, designed to perform fast molecular dynamics...
Understanding matter at the quantum scale of atoms and electrons is necessary for discovering promising new materials for energy transition or for the discovery of new medicines. We present a unique foundational model for quantum chemistry, designed to simulate matter at the atomic level, at the...
Developed by the OpenLLM-France community led by LINAGORA, Lucie-1B is a multilingual foundation language model with a focus on French language and culture. This project aims to offer a truly open alternative that complies with ethical and legal rules to the English-dominated generative AI landscape...
Faced with the explosion of astronomical data (hundreds of terabytes, hundreds of millions of objects), AION-1 unifies, for the first time, the analysis of heterogeneous observations in a single model of over 3 billion parameters. This foundational model, trained on Jean Zay, allows astrophysicists...
What if an autonomous car could learn simply by watching freely available driving videos online? Our project explores this radical idea: training a 1.2-billion-parameter model on 1,800 hours of unannotated dashcams, with no calibration or maps. This approach highlights the importance of scale and...
The vast majority of small open-source LLMs show significant performance variations depending on the language used. As part of this project, we propose a new version of the Vigogne model family: a series of SLMs specifically optimized for French. Vigogne offers enhanced capabilities for better...
Voltage-gated ion channels play a central role in cellular excitability. Mutations can induce pathologies such as cardiac arrhythmias, hypertension, or epilepsy. This project combines artificial intelligence and molecular dynamics simulations to reveal, at the atomic scale, the conformational...
Pleias 1.0 is the first family of language models trained exclusively on open data under a license that permits redistribution. Using the Jean Zay supercomputer, this project demonstrates that it is possible to reconcile performance, energy efficiency, and open science.