AI People Wiki

Builders & open source

Armand Joulin

Gemma Team (lead on the Gemma 2 and Gemma 3 reports; core contributor on the July 2026 Gemma 4 report), Google DeepMind

  • Gemma Team (lead on the Gemma 2 and Gemma 3 reports; core contributor on the July 2026 Gemma 4 report), Google DeepMind234

Armand Joulin is listed as the lead in the contributor lists of the Gemma 2 (2024) and Gemma 3 (2025) technical reports from Google DeepMind's Gemma Team, and among the leads of the first Gemma report (2024). The July 2026 Gemma 4 Technical Report from the same team lists Joulin among its core contributors and says the Gemma 4 models are released under an Apache 2.0 license. Earlier, Joulin was first author of Bag of Tricks for Efficient Text Classification (2016), written at Facebook AI Research, and of FastText.zip, two of the three papers listed in the references of the MIT-licensed fastText repository. Joulin is also a co-author of LLaMA: Open and Efficient Foundation Language Models (2023), written at Meta AI.1234567

At a glance

Current role
Gemma Team (lead on the Gemma 2 and Gemma 3 reports; core contributor on the July 2026 Gemma 4 report)
Sector
Company
Work listed
3
Sources cited
12
Profile reviewed

Research

Armand Joulin's work at Facebook AI Research included efficient text representation and classification. Joulin was first author of "Bag of Tricks for Efficient Text Classification" and "FastText.zip: Compressing text classification models" in 2016, and co-author with Piotr Bojanowski, Edouard Grave and Tomas Mikolov of "Enriching Word Vectors with Subword Information", the papers behind the fastText library.5896

Joulin also worked on self-supervised learning for computer vision. With Mathilde Caron and others, Joulin co-authored SwAV (2020), which learns visual features by contrasting cluster assignments, and DINO (2021), which studied the properties of self-supervised Vision Transformers. Joulin is also a co-author of the 2023 DINOv2 paper on learning visual features without supervision.101112

Open models

Joulin co-authored "LLaMA: Open and Efficient Foundation Language Models" (2023), written at Meta AI, whose abstract says the models were released to the research community. At Google DeepMind, Joulin appears among the leads of the first Gemma report in March 2024 and as the lead of the Gemma 2 (2024) and Gemma 3 (2025) technical reports.7123

The Gemma 4 Technical Report, submitted in July 2026, lists Joulin among its core contributors and says the Gemma 4 models are released under an Apache 2.0 license.4

Work and contributions

  • Gemma Open weights

    Listed among the leads of the Gemma (2024) report and as the lead of the Gemma 2 (2024) and Gemma 3 (2025) reports, and among the core contributors of the Gemma 4 Technical Report (2026); the Gemma 4 report says those models are released under an Apache 2.0 license.1234

  • LLaMA Published research

    Co-author of LLaMA: Open and Efficient Foundation Language Models (2023), whose author list gives a Meta AI affiliation and whose abstract says all models were released to the research community.7

  • fastText Open-source code

    First author of Bag of Tricks for Efficient Text Classification (2016) and of FastText.zip, and co-author of Enriching Word Vectors with Subword Information, the three papers listed in the references of the MIT-licensed fastText repository (archived in 2024).65

Timeline

  1. First author of "Bag of Tricks for Efficient Text Classification", the fastText classification paper.5
  2. Co-authored "Enriching Word Vectors with Subword Information" with Bojanowski, Grave and Mikolov.9
  3. First author of "FastText.zip", on compressing text classification models.8
  4. Co-authored SwAV, a self-supervised method that contrasts cluster assignments of image features.10
  5. Co-authored DINO, "Emerging Properties in Self-Supervised Vision Transformers".11
  6. Co-authored the LLaMA foundation language model paper at Meta AI.7
  7. Co-authored DINOv2, on learning robust visual features without supervision.12
  8. Listed among the leads of the first Gemma open-model report from Google DeepMind.1
  9. Listed as lead of the Gemma 2 technical report.2
  10. Listed as lead of the Gemma 3 technical report.3
  11. Listed among the core contributors of the Gemma 4 Technical Report.4

Other documented roles

Roles the sources document that are not confirmed as current.

Where they work

Suggest a correction or removal

Profiles cover public professional work only, and each statement cites a source. If something is out of date, missing or wrong, tell us and link a source. People can also ask for their profile to be corrected or removed.

References

12 sources · profile reviewed Oct 21, 2026 · data updated Oct 21, 2026

  1. Gemma: Open Models Based on Gemini Research and Technology (arXiv 2403.08295), contributor listarXiv · Paper · published Mar 13, 2024 · accessed Oct 21, 2026
  2. Gemma 2: Improving Open Language Models at a Practical Size (arXiv 2408.00118), contributor listarXiv · Paper · published Jul 31, 2024 · accessed Oct 21, 2026
  3. Gemma 3 Technical Report (arXiv 2503.19786), contributor listarXiv · Paper · published Mar 25, 2025 · accessed Oct 21, 2026
  4. Gemma 4 Technical Report (arXiv 2607.02770, v2 July 2026), core contributor list and license statementarXiv · Paper · published Jul 2, 2026 · accessed Oct 21, 2026
  5. Bag of Tricks for Efficient Text Classification (arXiv 1607.01759), Facebook AI Research affiliationarXiv · Paper · published Jul 6, 2016 · accessed Oct 21, 2026
  6. facebookresearch/fastText README (references and MIT license)facebookresearch (GitHub repository) · Repository · accessed Oct 21, 2026
  7. LLaMA: Open and Efficient Foundation Language Models (arXiv 2302.13971), author list and Meta AI affiliationarXiv · Paper · published Feb 27, 2023 · accessed Oct 21, 2026
  8. FastText.zip: Compressing text classification models (arXiv 1612.03651)arXiv · Paper · published Dec 12, 2016 · accessed Oct 21, 2026
  9. Enriching Word Vectors with Subword Information (arXiv 1607.04606)arXiv · Paper · published Jul 15, 2016 · accessed Oct 21, 2026
  10. Unsupervised Learning of Visual Features by Contrasting Cluster Assignments (arXiv 2006.09882)arXiv · Paper · published Jun 17, 2020 · accessed Oct 21, 2026
  11. Emerging Properties in Self-Supervised Vision Transformers (arXiv 2104.14294)arXiv · Paper · published Apr 29, 2021 · accessed Oct 21, 2026
  12. DINOv2: Learning Robust Visual Features without Supervision (arXiv 2304.07193)arXiv · Paper · published Apr 14, 2023 · accessed Oct 21, 2026