Timothée Lacroix

Timothée Lacroix graduated in computer science from the École normale supérieure in Paris in 2015 and took the Mathematics, Vision, Learning master's at Paris-Saclay. He joined Facebook AI Research in New York as an engineer that year and wrote his doctoral thesis there between 2016 and 2019, in collaboration with École des Ponts, on tensor factorization. The best-known result, a 2018 paper with Nicolas Usunier and Guillaume Obozinski, took the canonical tensor decomposition, a technique first described in 1927, and showed that with a new regularizer and a reformulated training objective it could match or beat newer methods for filling in missing facts in knowledge bases. He was also an author of PyTorch-BigGraph (2019), Facebook's open-source system for learning embeddings of graphs with billions of nodes and trillions of edges.

He stayed at Meta and worked with his longtime friend Guillaume Lample, including on HyperTree Proof Search, a 2022 neural theorem prover, and on LLaMA, whose February 2023 paper lists him among its authors. Weeks after LLaMA's release he and Lample left to found Mistral AI in Paris with Arthur Mensch; Mistral says the company was born in April 2023, and Lacroix became chief technology officer.

As CTO Lacroix leads Mistral's engineering teams, which run the training and serving systems behind its models, from the Apache 2.0 Mistral 7B of September 2023 to Mistral Large 3, trained on 3,000 Nvidia H200 GPUs and released with open weights in December 2025. In June 2025 Mistral launched Mistral Compute, an infrastructure business built on Nvidia reference designs that sells GPU capacity and a full AI stack to customers wanting an alternative to American and Chinese clouds. Mistral's 2026 AI Now Summit biography lists him as chief technical officer, leading the company's engineering teams.