Timothée Lacroix
Last reviewed
Sources
No citations yet
Review status
Needs citations
Revision
v2 · 1,945 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
No citations yet
Review status
Needs citations
Revision
v2 · 1,945 words
Add missing citations, update stale details, or suggest a clearer explanation.
Timothée Lacroix is a French artificial intelligence researcher and engineer who is a co-founder and the chief technology officer of Mistral AI, the Paris-based developer of open-weight large language models.[1][2] Before starting the company in 2023, he spent several years at Facebook AI Research (FAIR), later Meta AI, where he worked on machine learning for knowledge bases and was a co-author of the original LLaMA paper.[2][3][4] He founded Mistral AI together with Arthur Mensch and Guillaume Lample, two researchers he had worked alongside in Paris.[1][5]
Lacroix's work sits on the engineering side of large-scale machine learning, from his doctoral research on tensor methods for relational data to the systems and training pipelines behind frontier language models.[4][6] At Mistral AI he is responsible for the company's technical direction and its model research and infrastructure, complementing co-founder Guillaume Lample, who serves as chief science officer, and chief executive Arthur Mensch.[1][2] The three founders built Mistral into one of Europe's most prominent AI companies within a few years of its launch, reaching a valuation of about 11.7 billion euros by late 2025.[7][8]
Lacroix is a graduate of the École Normale Supérieure (ENS) in Paris.[5][9] He went on to complete a master's degree associated with the Paris region universities and then a PhD focused on machine learning, carried out in connection with the École des Ponts ParisTech (ENPC) while he was working at Facebook AI Research.[9][6] Unlike his two co-founders, Arthur Mensch and Guillaume Lample, who are both alumni of the École Polytechnique (class of 2011), Lacroix came through ENS rather than Polytechnique; the founders' paths converged through the Paris machine learning community and their shared time at Meta rather than a single school.[5][9]
His doctoral research centered on knowledge base completion, the problem of predicting missing facts in large relational datasets. As first author he published "Canonical Tensor Decomposition for Knowledge Base Completion" at the International Conference on Machine Learning (ICML) in 2018, work that reframed knowledge base completion as low-rank tensor decomposition and introduced a nuclear-norm-based regularizer that improved accuracy on standard benchmarks.[6][10] He later extended the approach to time-evolving data in "Tensor Decompositions for Temporal Knowledge Base Completion," presented at the International Conference on Learning Representations (ICLR) in 2020.[10][11] His doctoral dissertation, "Tensor decompositions for knowledge base completion," was completed through the University of Paris-Est and gathers this line of work, which remains among his most cited research, with the 2018 ICML paper accumulating roughly 560 citations and the 2020 ICLR paper roughly 490 by 2026.[4][15]
Lacroix joined Facebook AI Research in Paris, first as a PhD student and subsequently as a research engineer, and remained there until co-founding Mistral.[5][9] His earliest work at the company continued his thesis line on embedding methods and tensor factorization for knowledge bases, an area where representations of entities and relations are learned so that plausible facts score higher than implausible ones.[6][10]
His work at FAIR ranged across several research lines beyond knowledge bases. In 2016 he was a co-author of "TorchCraft," a library that connected machine learning code to the real-time strategy game StarCraft so that agents could be trained and tested in that environment.[15][16] He later co-authored "PyTorch-BigGraph: A Large Scale Graph Embedding System," presented at the SysML (later MLSys) conference in 2019, which described a system for learning embeddings of graphs with billions of nodes and edges by partitioning the graph across machines; the paper has been cited several hundred times and the accompanying software was released as open source.[4][15][17] Lacroix also contributed to FAIR's work on automated mathematical reasoning. He is a co-author of "HyperTree Proof Search for Neural Theorem Proving," published at the Conference on Neural Information Processing Systems (NeurIPS) in 2022 and led by Guillaume Lample, which introduced a search algorithm inspired by AlphaZero that proved 82.6 percent of a held-out set of Metamath theorems after online training and raised the state of the art on the Lean miniF2F benchmark from 31 to 42 percent.[18][15] A related collaboration, "Draft, Sketch, and Prove," which guided formal theorem provers with informal proofs, appeared at ICLR in 2023.[15]
His best-known contribution at Meta came from large language models. He is listed as a co-author of "LLaMA: Open and Efficient Foundation Language Models," the February 2023 paper that introduced Meta's first LLaMA family of models and was led by Guillaume Lample, with Hugo Touvron as lead author.[3] LLaMA showed that comparatively small models trained on more tokens could match or exceed much larger systems, and its release, including the leak and subsequent open distribution of the weights, helped seed a wave of open-source language model development.[3] By 2026 the LLaMA paper had become by far Lacroix's most cited work, with on the order of 28,000 citations, and it dominates the roughly 35,000 total citations recorded on his Google Scholar profile.[4] French press covering Mistral has described Lacroix as a former Meta researcher and Lample as one of LLaMA's creators, the engineering and research background the pair carried into their own company.[5]
Lacroix co-founded Mistral AI on 28 April 2023 with Arthur Mensch, who had spent about three years at Google DeepMind, and Guillaume Lample, his former FAIR colleague.[1][5] The company positioned itself as a European builder of open and efficient foundation models, an explicit contrast to the more closed approach of larger United States labs.[7]
Mistral attracted capital unusually quickly for a company at its stage. The table below summarizes its main early funding milestones.
| Round | Date | Amount | Notes |
|---|---|---|---|
| Seed / Series A | June 2023 | about 105 million euros | Led by Lightspeed Venture Partners; valuation around 240 million euros[7][12] |
| Series B | December 2023 | about 385 million euros | Investors including Andreessen Horowitz; valuation near 2 billion dollars[7][12] |
| Series C | September 2025 | about 1.7 billion euros | Led by ASML, with Nvidia among backers; valuation about 11.7 billion euros[8][13] |
The 2025 round, which valued Mistral at roughly 11.7 billion euros (about 14 billion dollars), was led by the Dutch semiconductor-equipment maker ASML and made the three founders among the first home-grown AI billionaires in France on paper.[8][13] According to the Bloomberg Billionaires Index, the September 2025 round left Lacroix, then 34, with a net worth of about 1.1 billion dollars, the same figure attributed to each of his two co-founders, who together hold stakes reported at no less than 8 percent each in the company.[19][20] Mistral's products include its open-weight model families and Le Chat, a conversational assistant marketed as a European alternative to other mainstream chatbots.[7][8]
As CTO, Lacroix leads Mistral's technical organization, overseeing model research, training, and the engineering of the systems that serve the company's models.[1][2] Public profiles and the company's own about page list him simply as "Co-founder and CTO," alongside Mensch as CEO and Lample as chief science officer.[1][2] In interviews and talks he has emphasized efficiency in model training and inference, the discipline of getting strong results from constrained compute budgets, which has been central to Mistral's strategy of competing with far larger rivals.[14] Compared with the more public-facing Mensch, Lacroix has kept a comparatively low profile in the press while concentrating on the company's engineering.[20]
That emphasis on efficiency is visible in the models he has helped ship at Mistral. He is a co-author of the technical report for Mistral 7B, the company's first model, released in September 2023 under the permissive Apache 2.0 license; the 7.3-billion-parameter model used grouped-query attention and sliding window attention to lower memory and inference cost while outperforming the larger Llama 2 13B on many benchmarks.[21][22] He is also credited on the report for Mixtral of Experts, the sparse mixture-of-experts model that followed, which by 2026 had become his second most cited paper, and on Pixtral 12B, Mistral's multimodal model that added image understanding.[4][15][23]
His Google Scholar profile, which lists affiliations with Facebook AI Research and the École des Ponts, records his research output on knowledge base completion and language modeling, including the LLaMA paper among his most cited work.[4]
| Year | Work | Venue | Role |
|---|---|---|---|
| 2018 | Canonical Tensor Decomposition for Knowledge Base Completion[6] | ICML | First author |
| 2019 | PyTorch-BigGraph: A Large Scale Graph Embedding System[17] | SysML / MLSys | Co-author |
| 2020 | Tensor Decompositions for Temporal Knowledge Base Completion[10] | ICLR | First author |
| 2022 | HyperTree Proof Search for Neural Theorem Proving[18] | NeurIPS | Co-author |
| 2023 | LLaMA: Open and Efficient Foundation Language Models[3] | preprint (Meta AI) | Co-author |
| 2023 | Mistral 7B[21] | preprint (Mistral AI) | Co-author |
| 2024 | Mixtral of Experts[23] | preprint (Mistral AI) | Co-author |