Enregistré dans:
| Auteurs principaux: | , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Publié: |
2025
|
| Sujets: | |
| Accès en ligne: | https://arxiv.org/abs/2506.10910 |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866909647103852544 |
|---|---|
| author | Mistral-AI : Rastogi, Abhinav Jiang, Albert Q. Lo, Andy Berrada, Gabrielle Lample, Guillaume Rute, Jason Barmentlo, Joep Yadav, Karmesh Khandelwal, Kartik Chandu, Khyathi Raghavi Blier, Léonard Saulnier, Lucile Dinot, Matthieu Darrin, Maxime Gupta, Neha Soletskyi, Roman Vaze, Sagar Scao, Teven Le Wang, Yihan Yang, Adam Liu, Alexander H. Sablayrolles, Alexandre Héliou, Amélie Martin, Amélie Ehrenberg, Andy Agarwal, Anmol Roux, Antoine Darcet, Arthur Mensch, Arthur Bout, Baptiste Rozière, Baptiste De Monicault, Baudouin Bamford, Chris Wallenwein, Christian Renaudin, Christophe Lanfranchi, Clémence Dabert, Darius Mizelle, Devon Casas, Diego de las Chane-Sane, Elliot Fugier, Emilien Hanna, Emma Bou Delerce, Gauthier Guinet, Gauthier Novikov, Georgii Martin, Guillaume Jaju, Himanshu Ludziejewski, Jan Chabran, Jean-Hadrien Delignon, Jean-Malo Studnia, Joachim Amar, Jonas Roberts, Josselin Somerville Denize, Julien Saxena, Karan Jain, Kush Zhao, Lingxiao Martin, Louis Gao, Luyu Lavaud, Lélio Renard Pellat, Marie Guillaumin, Mathilde Felardos, Mathis Augustin, Maximilian Seznec, Mickaël Raghuraman, Nikhil Duchenne, Olivier Wang, Patricia von Platen, Patrick Saffer, Patryk Jacob, Paul Wambergue, Paul Kurylowicz, Paula Muddireddy, Pavankumar Reddy Chagniot, Philomène Stock, Pierre Agrawal, Pravesh Sauvestre, Romain Delacourt, Rémi Gandhi, Sanchit Subramanian, Sandeep Dalal, Shashwat Gandhi, Siddharth Ghosh, Soham Mishra, Srijan Aithal, Sumukh Antoniak, Szymon Schueller, Thibault Lavril, Thibaut Robert, Thomas Wang, Thomas Lacroix, Timothée Nemychnikova, Valeriia Paltz, Victor Richard, Virgile Li, Wen-Ding Marshall, William Zhang, Xuanyu Tang, Yunhao |
| author_facet | Mistral-AI : Rastogi, Abhinav Jiang, Albert Q. Lo, Andy Berrada, Gabrielle Lample, Guillaume Rute, Jason Barmentlo, Joep Yadav, Karmesh Khandelwal, Kartik Chandu, Khyathi Raghavi Blier, Léonard Saulnier, Lucile Dinot, Matthieu Darrin, Maxime Gupta, Neha Soletskyi, Roman Vaze, Sagar Scao, Teven Le Wang, Yihan Yang, Adam Liu, Alexander H. Sablayrolles, Alexandre Héliou, Amélie Martin, Amélie Ehrenberg, Andy Agarwal, Anmol Roux, Antoine Darcet, Arthur Mensch, Arthur Bout, Baptiste Rozière, Baptiste De Monicault, Baudouin Bamford, Chris Wallenwein, Christian Renaudin, Christophe Lanfranchi, Clémence Dabert, Darius Mizelle, Devon Casas, Diego de las Chane-Sane, Elliot Fugier, Emilien Hanna, Emma Bou Delerce, Gauthier Guinet, Gauthier Novikov, Georgii Martin, Guillaume Jaju, Himanshu Ludziejewski, Jan Chabran, Jean-Hadrien Delignon, Jean-Malo Studnia, Joachim Amar, Jonas Roberts, Josselin Somerville Denize, Julien Saxena, Karan Jain, Kush Zhao, Lingxiao Martin, Louis Gao, Luyu Lavaud, Lélio Renard Pellat, Marie Guillaumin, Mathilde Felardos, Mathis Augustin, Maximilian Seznec, Mickaël Raghuraman, Nikhil Duchenne, Olivier Wang, Patricia von Platen, Patrick Saffer, Patryk Jacob, Paul Wambergue, Paul Kurylowicz, Paula Muddireddy, Pavankumar Reddy Chagniot, Philomène Stock, Pierre Agrawal, Pravesh Sauvestre, Romain Delacourt, Rémi Gandhi, Sanchit Subramanian, Sandeep Dalal, Shashwat Gandhi, Siddharth Ghosh, Soham Mishra, Srijan Aithal, Sumukh Antoniak, Szymon Schueller, Thibault Lavril, Thibaut Robert, Thomas Wang, Thomas Lacroix, Timothée Nemychnikova, Valeriia Paltz, Victor Richard, Virgile Li, Wen-Ding Marshall, William Zhang, Xuanyu Tang, Yunhao |
| contents | We introduce Magistral, Mistral's first reasoning model and our own scalable reinforcement learning (RL) pipeline. Instead of relying on existing implementations and RL traces distilled from prior models, we follow a ground up approach, relying solely on our own models and infrastructure. Notably, we demonstrate a stack that enabled us to explore the limits of pure RL training of LLMs, present a simple method to force the reasoning language of the model, and show that RL on text data alone maintains most of the initial checkpoint's capabilities. We find that RL on text maintains or improves multimodal understanding, instruction following and function calling. We present Magistral Medium, trained for reasoning on top of Mistral Medium 3 with RL alone, and we open-source Magistral Small (Apache 2.0) which further includes cold-start data from Magistral Medium. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2506_10910 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Magistral Mistral-AI : Rastogi, Abhinav Jiang, Albert Q. Lo, Andy Berrada, Gabrielle Lample, Guillaume Rute, Jason Barmentlo, Joep Yadav, Karmesh Khandelwal, Kartik Chandu, Khyathi Raghavi Blier, Léonard Saulnier, Lucile Dinot, Matthieu Darrin, Maxime Gupta, Neha Soletskyi, Roman Vaze, Sagar Scao, Teven Le Wang, Yihan Yang, Adam Liu, Alexander H. Sablayrolles, Alexandre Héliou, Amélie Martin, Amélie Ehrenberg, Andy Agarwal, Anmol Roux, Antoine Darcet, Arthur Mensch, Arthur Bout, Baptiste Rozière, Baptiste De Monicault, Baudouin Bamford, Chris Wallenwein, Christian Renaudin, Christophe Lanfranchi, Clémence Dabert, Darius Mizelle, Devon Casas, Diego de las Chane-Sane, Elliot Fugier, Emilien Hanna, Emma Bou Delerce, Gauthier Guinet, Gauthier Novikov, Georgii Martin, Guillaume Jaju, Himanshu Ludziejewski, Jan Chabran, Jean-Hadrien Delignon, Jean-Malo Studnia, Joachim Amar, Jonas Roberts, Josselin Somerville Denize, Julien Saxena, Karan Jain, Kush Zhao, Lingxiao Martin, Louis Gao, Luyu Lavaud, Lélio Renard Pellat, Marie Guillaumin, Mathilde Felardos, Mathis Augustin, Maximilian Seznec, Mickaël Raghuraman, Nikhil Duchenne, Olivier Wang, Patricia von Platen, Patrick Saffer, Patryk Jacob, Paul Wambergue, Paul Kurylowicz, Paula Muddireddy, Pavankumar Reddy Chagniot, Philomène Stock, Pierre Agrawal, Pravesh Sauvestre, Romain Delacourt, Rémi Gandhi, Sanchit Subramanian, Sandeep Dalal, Shashwat Gandhi, Siddharth Ghosh, Soham Mishra, Srijan Aithal, Sumukh Antoniak, Szymon Schueller, Thibault Lavril, Thibaut Robert, Thomas Wang, Thomas Lacroix, Timothée Nemychnikova, Valeriia Paltz, Victor Richard, Virgile Li, Wen-Ding Marshall, William Zhang, Xuanyu Tang, Yunhao Computation and Language We introduce Magistral, Mistral's first reasoning model and our own scalable reinforcement learning (RL) pipeline. Instead of relying on existing implementations and RL traces distilled from prior models, we follow a ground up approach, relying solely on our own models and infrastructure. Notably, we demonstrate a stack that enabled us to explore the limits of pure RL training of LLMs, present a simple method to force the reasoning language of the model, and show that RL on text data alone maintains most of the initial checkpoint's capabilities. We find that RL on text maintains or improves multimodal understanding, instruction following and function calling. We present Magistral Medium, trained for reasoning on top of Mistral Medium 3 with RL alone, and we open-source Magistral Small (Apache 2.0) which further includes cold-start data from Magistral Medium. |
| title | Magistral |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2506.10910 |