Guillaume Lample
Last reviewed
Jun 5, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v2 · 2,275 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 5, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v2 · 2,275 words
Add missing citations, update stale details, or suggest a clearer explanation.
Guillaume Lample (born 8 October 1990) is a French artificial intelligence researcher and entrepreneur, best known as a senior author of the LLaMA paper produced at Meta AI and as a co-founder and chief scientist of Mistral AI.[1][2][3] Before founding Mistral in 2023, he spent roughly seven years as a research scientist at Meta's Fundamental AI Research lab (FAIR) in Paris, where his work spanned machine translation, multilingual representation learning, game-playing agents, symbolic mathematics, and the open foundation models that became LLaMA.[4][5]
Lample's research career has moved through several distinct lines of work, most of it carried out at FAIR Paris between 2016 and early 2023. His early publications addressed sequence labeling and named entity recognition, after which he became a central contributor to a body of work on unsupervised and cross-lingual machine translation, word embedding alignment, and multilingual pretraining.[5][6] He later worked on applying deep learning to symbolic mathematics and automated theorem proving, and was a senior author on the first LLaMA release, a family of open large language models that Meta published in February 2023.[1][4] Within weeks of that release he left Meta to help start Mistral AI, where he serves as chief scientist and leads model research.[3][7]
His Google Scholar profile, which lists Mistral AI as his affiliation, records on the order of 66,000 citations and an h-index in the low 40s, reflecting the wide adoption of LLaMA, the Mistral and Mixtral models, and his earlier translation and representation-learning papers.[2]
Lample was born on 8 October 1990 in Brest, in the Brittany region of north-western France, and the French regional press has described him as a Breton "prodigy" of artificial intelligence.[16][17] His birth date and birthplace are recorded in the Wikidata authority entry on him, which cites reporting in the Breton daily Le Telegramme, and the French national authority file (IdRef) likewise lists his year of birth as 1990.[16][17]
He is a graduate of the École Polytechnique in France, where he was a member of the 2011 class and where he later met the two colleagues with whom he would found Mistral AI.[7][8][16] He went on to earn a Master of Language Technologies at the Language Technologies Institute of Carnegie Mellon University, completing the degree in 2017.[8][18] His graduate work in natural language processing was supervised by faculty including Chris Dyer, and his subsequent collaboration with Dyer is reflected in the 2016 paper on neural architectures for named entity recognition, one of his most cited works.[2][5]
He completed a PhD in computer science under a French CIFRE arrangement, pursued jointly at FAIR Paris and at Sorbonne Université (formerly Université Pierre et Marie Curie), in the Machine Learning and Information Access (MLIA) team of the LIP6 laboratory.[4][9] His thesis, titled "Unsupervised Machine Translation," was defended on 17 October 2019 and was supervised by Ludovic Denoyer and Antoine Bordes.[9][17] The thesis consolidated his FAIR research on building translation systems from monolingual data alone, without parallel corpora.[9]
Lample joined Facebook AI Research (later Meta AI) in Paris in 2016 and remained there as a research scientist until early 2023.[4][7] His GitHub profile from that period described his focus as "unsupervised machine translation, symbolic mathematics and theorem proving," and he authored or co-authored several widely used research repositories, including the MUSE library for multilingual word embeddings and the Arnold agent for the video game DOOM.[10]
His most prominent contribution at Meta was the LLaMA project. The paper "LLaMA: Open and Efficient Foundation Language Models," submitted on 27 February 2023, introduced a collection of foundation models ranging from 7 billion to 65 billion parameters, trained only on publicly available data and released to the research community.[1] The work had fourteen authors, with Hugo Touvron listed first and Lample listed last, the position that by convention in machine learning denotes the senior author who guided the project.[1][2] As of 2026 his Google Scholar profile credits the LLaMA paper with more than 28,000 citations, by a wide margin his most cited work.[2] LLaMA's release was influential in catalyzing the open-weights ecosystem, and the smaller models in particular demonstrated that comparatively compact networks trained on more tokens could rival much larger systems.[1]
LLaMA capped a productive run at FAIR. Lample's colleague Timothée Lacroix, later a co-founder of Mistral, was also among the LLaMA authors, and the two left Meta together in early 2023.[3][11]
A substantial part of Lample's FAIR work concerned learning to translate with little or no parallel data. The 2018 paper "Word translation without parallel data" introduced methods for aligning word embedding spaces across languages without supervision, and "Unsupervised Machine Translation Using Monolingual Corpora Only" extended the idea to full sentence translation.[2][6] He was also the first author, with Alexis Conneau, of the cross-lingual language model pretraining work that introduced XLM, an approach for learning multilingual representations later extended in the XLM-R line, and he co-authored the XNLI benchmark for evaluating cross-lingual sentence understanding.[2][19] His Google Scholar metrics place the named entity recognition paper above 6,000 citations and the XLM pretraining paper around 4,000, underlining the lasting influence of this line of work.[2] These contributions sit within his broader interest in multilingual NLP and representation learning.[5]
Earlier in his career Lample worked on deep reinforcement learning for video games. With his Carnegie Mellon classmate Devendra Singh Chaplot he built Arnold, an agent that played the first-person shooter DOOM using only the pixels visible on screen, combining a deep Q-network for navigation with a deep recurrent Q-network for combat.[20] Competing as the team "The Terminators," Arnold placed second in both tracks of the ViZDoom competition held at the IEEE Conference on Computational Intelligence and Games in September 2016, recording the lowest number of deaths and the best kill-to-death ratio of the field, and the open-source release of the agent went on to win the 2017 edition of the ViZDoom AI Competition.[20][21] The associated paper, "Playing FPS Games with Deep Reinforcement Learning," was presented at AAAI in 2017.[2][10]
Lample also helped open up the application of neural sequence models to mathematics. With François Charton he published "Deep Learning for Symbolic Mathematics" at ICLR 2020, showing that transformer models could perform symbolic integration and solve differential equations by treating mathematics as a translation problem.[12] He continued in this direction with work on automated theorem proving, including "HyperTree Proof Search for Neural Theorem Proving," presented at NeurIPS in 2022, which paired a learned policy with a search procedure inspired by AlphaZero to construct formal proofs.[13]
| Year | Work | Venue | Role |
|---|---|---|---|
| 2016 | Neural Architectures for Named Entity Recognition | NAACL | First author |
| 2017 | Playing FPS Games with Deep Reinforcement Learning | AAAI | First author |
| 2018 | Word Translation Without Parallel Data | ICLR | First author |
| 2018 | Unsupervised Machine Translation Using Monolingual Corpora Only | ICLR | First author |
| 2019 | Cross-lingual Language Model Pretraining (XLM) | NeurIPS | First author |
| 2020 | Deep Learning for Symbolic Mathematics | ICLR | First author |
| 2022 | HyperTree Proof Search for Neural Theorem Proving | NeurIPS | First author |
| 2023 | LLaMA: Open and Efficient Foundation Language Models | arXiv | Senior author |
| 2023 | Mistral 7B | arXiv | Co-author |
| 2024 | Mixtral of Experts | arXiv | Co-author |
| 2025 | Magistral | arXiv | Co-author |
Citation figures and authorship positions are drawn from his Google Scholar profile and the original papers.[1][2]
Lample co-founded Mistral AI in Paris in 2023 with Arthur Mensch, who became chief executive, and Timothée Lacroix, who became chief technology officer.[3][11] The three founders had met during their studies at the École Polytechnique; Lample and Lacroix came from Meta, while Mensch had worked at Google DeepMind.[3] The company was incorporated on 28 April 2023 and positioned itself as a European developer of open and efficient foundation models, an emphasis consistent with Lample's open-release work on LLaMA.[3]
Mistral attracted large early investment, raising about 105 million euros in a June 2023 seed round and a 385-million-euro round in December 2023 that lifted its valuation past 2 billion euros.[3] The company released its first model, Mistral 7B, in September 2023, followed by the Mixtral family built on a sparse mixture of experts architecture, and subsequent model generations including the proprietary Mistral Large line.[2][3] Many of these systems were distributed under open weights licenses, reflecting the founders' stated commitment to openness.[3]
Lample serves as Mistral AI's co-founder and chief scientist, a role Mistral's own materials describe as "Co-founder and Chief Science Officer," with responsibility for the company's model research and training.[7][14] He has framed Mistral's strategy in terms of building its training stack independently, telling reporters that the company "did everything from scratch, basically because we wanted to learn the expertise we have."[22]
In June 2025 he was a co-author on Magistral, described in the accompanying technical report as Mistral's first reasoning model, released in an open-weights Magistral Small version under the Apache 2.0 license alongside a larger Magistral Medium served through Mistral's API.[22][23] A distinguishing feature emphasized at launch was that the model exposes its full chain of thought to the user in the user's own language rather than defaulting to English.[22] In December 2025 Mistral released the Mistral 3 generation, which included a set of small dense models and Mistral Large 3, a sparse mixture-of-experts model with roughly 41 billion active and 675 billion total parameters trained on NVIDIA H200 GPUs and published under the Apache 2.0 license.[24][25] In March 2026 the company raised about 830 million US dollars in debt financing from a syndicate of banks to build a data center near Paris, part of a broader European infrastructure expansion.[26]
In September 2025 a funding round that valued Mistral at roughly 11.7 billion euros (about 13.8 billion US dollars) made Lample, Mensch, and Lacroix the first artificial intelligence billionaires in France, each holding a stake of at least 8 percent in the company, according to the Bloomberg Billionaires Index.[14][15] The round, Mistral's Series C, raised about 1.7 billion euros and was led by the Dutch chip-equipment maker ASML, which became the company's largest shareholder with a stake of about 11 percent.[14][15][27] Mistral's valuation continued to rise in subsequent reporting, and the company is widely regarded as Europe's most prominent independent artificial intelligence developer.[3][27]