Copertina

Llama-Nemotron: Efficient Reasoning Models

Fuente: arXiv

Salvato in:

Dettagli Bibliografici
Autori principali:	Bercovich, Akhiad, Levy, Itay, Golan, Izik, Dabbah, Mohammad, El-Yaniv, Ran, Puny, Omri, Galil, Ido, Moshe, Zach, Ronen, Tomer, Nabwani, Najeeb, Shahaf, Ido, Tropp, Oren, Karpas, Ehud, Zilberstein, Ran, Zeng, Jiaqi, Singhal, Soumye, Bukharin, Alexander, Zhang, Yian, Konuk, Tugrul, Shen, Gerald, Mahabaleshwarkar, Ameya Sunil, Kartal, Bilal, Suhara, Yoshi, Delalleau, Olivier, Chen, Zijia, Wang, Zhilin, Mosallanezhad, David, Renduchintala, Adi, Qian, Haifeng, Rekesh, Dima, Jia, Fei, Majumdar, Somshubra, Noroozi, Vahid, Ahmad, Wasi Uddin, Narenthiran, Sean, Ficek, Aleksander, Samadi, Mehrzad, Huang, Jocelyn, Jain, Siddhartha, Gitman, Igor, Moshkov, Ivan, Du, Wei, Toshniwal, Shubham, Armstrong, George, Kisacanin, Branislav, Novikov, Matvei, Gitman, Daria, Bakhturina, Evelina, Varshney, Prasoon, Narsimhan, Makesh, Scowcroft, Jane Polak, Kamalu, John, Su, Dan, Kong, Kezhi, Kliegl, Markus, Mahabadi, Rabeeh Karimi, Lin, Ying, Satheesh, Sanjeev, Parmar, Jupinder, Gundecha, Pritam, Norick, Brandon, Jennings, Joseph, Prabhumoye, Shrimai, Akter, Syeda Nahida, Patwary, Mostofa, Khattar, Abhinav, Narayanan, Deepak, Waleffe, Roger, Zhang, Jimmy, Su, Bor-Yiing, Huang, Guyue, Kong, Terry, Chadha, Parth, Jain, Sahil, Harvey, Christine, Segal, Elad, Huang, Jining, Kashirsky, Sergey, McQueen, Robert, Putterman, Izzy, Lam, George, Venkatesan, Arun, Wu, Sherry, Nguyen, Vinh, Kilaru, Manoj, Wang, Andrew, Warno, Anna, Somasamudramath, Abhilash, Bhaskar, Sandip, Dong, Maka, Assaf, Nave, Mor, Shahar, Argov, Omer Ullman, Junkin, Scot, Romanenko, Oleksandr, Larroy, Pedro, Katariya, Monika, Rovinelli, Marco, Balas, Viji, Edelman, Nicholas, Bhiwandiwalla, Anahita, Subramaniam, Muthu, Ithape, Smita, Ramamoorthy, Karthik, Wu, Yuting, Velury, Suguna Varshini, Almog, Omri, Daw, Joyjit, Fridman, Denys, Galinkin, Erick, Evans, Michael, Ghosh, Shaona, Luna, Katherine, Derczynski, Leon, Pope, Nikki, Long, Eileen, Schneider, Seth, Siman, Guillermo, Grzegorzek, Tomasz, Ribalta, Pablo, Alexiuk, Chris, Conway, Joey, Saar, Trisha, Guan, Ann, Pawelec, Krzysztof, Prayaga, Shyamala, Kuchaiev, Oleksii, Ginsburg, Boris, Olabiyi, Oluwatobi, Briski, Kari, Cohen, Jonathan, Catanzaro, Bryan, Alben, Jonah, Geifman, Yonatan, Chung, Eric
Natura:	Preprint
Pubblicazione:	2025
Soggetti:	Computation and Language Artificial Intelligence Machine Learning
Accesso online:	Acceder al recurso
Tags:	Aggiungi Tag Nessun Tag, puoi essere il primo ad aggiungerne!!

_version_	1866915484994109440
author	Bercovich, Akhiad Levy, Itay Golan, Izik Dabbah, Mohammad El-Yaniv, Ran Puny, Omri Galil, Ido Moshe, Zach Ronen, Tomer Nabwani, Najeeb Shahaf, Ido Tropp, Oren Karpas, Ehud Zilberstein, Ran Zeng, Jiaqi Singhal, Soumye Bukharin, Alexander Zhang, Yian Konuk, Tugrul Shen, Gerald Mahabaleshwarkar, Ameya Sunil Kartal, Bilal Suhara, Yoshi Delalleau, Olivier Chen, Zijia Wang, Zhilin Mosallanezhad, David Renduchintala, Adi Qian, Haifeng Rekesh, Dima Jia, Fei Majumdar, Somshubra Noroozi, Vahid Ahmad, Wasi Uddin Narenthiran, Sean Ficek, Aleksander Samadi, Mehrzad Huang, Jocelyn Jain, Siddhartha Gitman, Igor Moshkov, Ivan Du, Wei Toshniwal, Shubham Armstrong, George Kisacanin, Branislav Novikov, Matvei Gitman, Daria Bakhturina, Evelina Varshney, Prasoon Narsimhan, Makesh Scowcroft, Jane Polak Kamalu, John Su, Dan Kong, Kezhi Kliegl, Markus Mahabadi, Rabeeh Karimi Lin, Ying Satheesh, Sanjeev Parmar, Jupinder Gundecha, Pritam Norick, Brandon Jennings, Joseph Prabhumoye, Shrimai Akter, Syeda Nahida Patwary, Mostofa Khattar, Abhinav Narayanan, Deepak Waleffe, Roger Zhang, Jimmy Su, Bor-Yiing Huang, Guyue Kong, Terry Chadha, Parth Jain, Sahil Harvey, Christine Segal, Elad Huang, Jining Kashirsky, Sergey McQueen, Robert Putterman, Izzy Lam, George Venkatesan, Arun Wu, Sherry Nguyen, Vinh Kilaru, Manoj Wang, Andrew Warno, Anna Somasamudramath, Abhilash Bhaskar, Sandip Dong, Maka Assaf, Nave Mor, Shahar Argov, Omer Ullman Junkin, Scot Romanenko, Oleksandr Larroy, Pedro Katariya, Monika Rovinelli, Marco Balas, Viji Edelman, Nicholas Bhiwandiwalla, Anahita Subramaniam, Muthu Ithape, Smita Ramamoorthy, Karthik Wu, Yuting Velury, Suguna Varshini Almog, Omri Daw, Joyjit Fridman, Denys Galinkin, Erick Evans, Michael Ghosh, Shaona Luna, Katherine Derczynski, Leon Pope, Nikki Long, Eileen Schneider, Seth Siman, Guillermo Grzegorzek, Tomasz Ribalta, Pablo Katariya, Monika Alexiuk, Chris Conway, Joey Saar, Trisha Guan, Ann Pawelec, Krzysztof Prayaga, Shyamala Kuchaiev, Oleksii Ginsburg, Boris Olabiyi, Oluwatobi Briski, Kari Cohen, Jonathan Catanzaro, Bryan Alben, Jonah Geifman, Yonatan Chung, Eric
author_facet	Bercovich, Akhiad Levy, Itay Golan, Izik Dabbah, Mohammad El-Yaniv, Ran Puny, Omri Galil, Ido Moshe, Zach Ronen, Tomer Nabwani, Najeeb Shahaf, Ido Tropp, Oren Karpas, Ehud Zilberstein, Ran Zeng, Jiaqi Singhal, Soumye Bukharin, Alexander Zhang, Yian Konuk, Tugrul Shen, Gerald Mahabaleshwarkar, Ameya Sunil Kartal, Bilal Suhara, Yoshi Delalleau, Olivier Chen, Zijia Wang, Zhilin Mosallanezhad, David Renduchintala, Adi Qian, Haifeng Rekesh, Dima Jia, Fei Majumdar, Somshubra Noroozi, Vahid Ahmad, Wasi Uddin Narenthiran, Sean Ficek, Aleksander Samadi, Mehrzad Huang, Jocelyn Jain, Siddhartha Gitman, Igor Moshkov, Ivan Du, Wei Toshniwal, Shubham Armstrong, George Kisacanin, Branislav Novikov, Matvei Gitman, Daria Bakhturina, Evelina Varshney, Prasoon Narsimhan, Makesh Scowcroft, Jane Polak Kamalu, John Su, Dan Kong, Kezhi Kliegl, Markus Mahabadi, Rabeeh Karimi Lin, Ying Satheesh, Sanjeev Parmar, Jupinder Gundecha, Pritam Norick, Brandon Jennings, Joseph Prabhumoye, Shrimai Akter, Syeda Nahida Patwary, Mostofa Khattar, Abhinav Narayanan, Deepak Waleffe, Roger Zhang, Jimmy Su, Bor-Yiing Huang, Guyue Kong, Terry Chadha, Parth Jain, Sahil Harvey, Christine Segal, Elad Huang, Jining Kashirsky, Sergey McQueen, Robert Putterman, Izzy Lam, George Venkatesan, Arun Wu, Sherry Nguyen, Vinh Kilaru, Manoj Wang, Andrew Warno, Anna Somasamudramath, Abhilash Bhaskar, Sandip Dong, Maka Assaf, Nave Mor, Shahar Argov, Omer Ullman Junkin, Scot Romanenko, Oleksandr Larroy, Pedro Katariya, Monika Rovinelli, Marco Balas, Viji Edelman, Nicholas Bhiwandiwalla, Anahita Subramaniam, Muthu Ithape, Smita Ramamoorthy, Karthik Wu, Yuting Velury, Suguna Varshini Almog, Omri Daw, Joyjit Fridman, Denys Galinkin, Erick Evans, Michael Ghosh, Shaona Luna, Katherine Derczynski, Leon Pope, Nikki Long, Eileen Schneider, Seth Siman, Guillermo Grzegorzek, Tomasz Ribalta, Pablo Katariya, Monika Alexiuk, Chris Conway, Joey Saar, Trisha Guan, Ann Pawelec, Krzysztof Prayaga, Shyamala Kuchaiev, Oleksii Ginsburg, Boris Olabiyi, Oluwatobi Briski, Kari Cohen, Jonathan Catanzaro, Bryan Alben, Jonah Geifman, Yonatan Chung, Eric
contents	We introduce the Llama-Nemotron series of models, an open family of heterogeneous reasoning models that deliver exceptional reasoning capabilities, inference efficiency, and an open license for enterprise use. The family comes in three sizes -- Nano (8B), Super (49B), and Ultra (253B) -- and performs competitively with state-of-the-art reasoning models such as DeepSeek-R1 while offering superior inference throughput and memory efficiency. In this report, we discuss the training procedure for these models, which entails using neural architecture search from Llama 3 models for accelerated inference, knowledge distillation, and continued pretraining, followed by a reasoning-focused post-training stage consisting of two main parts: supervised fine-tuning and large scale reinforcement learning. Llama-Nemotron models are the first open-source models to support a dynamic reasoning toggle, allowing users to switch between standard chat and reasoning modes during inference. To further support open research and facilitate model development, we provide the following resources: 1. We release the Llama-Nemotron reasoning models -- LN-Nano, LN-Super, and LN-Ultra -- under the commercially permissive NVIDIA Open Model License Agreement. 2. We release the complete post-training dataset: Llama-Nemotron-Post-Training-Dataset. 3. We also release our training codebases: NeMo, NeMo-Aligner, and Megatron-LM.
format	Preprint
id	arxiv_https___arxiv_org_abs_2505_00949
institution	arXiv
publishDate	2025
record_format	arxiv
spellingShingle	Llama-Nemotron: Efficient Reasoning Models Bercovich, Akhiad Levy, Itay Golan, Izik Dabbah, Mohammad El-Yaniv, Ran Puny, Omri Galil, Ido Moshe, Zach Ronen, Tomer Nabwani, Najeeb Shahaf, Ido Tropp, Oren Karpas, Ehud Zilberstein, Ran Zeng, Jiaqi Singhal, Soumye Bukharin, Alexander Zhang, Yian Konuk, Tugrul Shen, Gerald Mahabaleshwarkar, Ameya Sunil Kartal, Bilal Suhara, Yoshi Delalleau, Olivier Chen, Zijia Wang, Zhilin Mosallanezhad, David Renduchintala, Adi Qian, Haifeng Rekesh, Dima Jia, Fei Majumdar, Somshubra Noroozi, Vahid Ahmad, Wasi Uddin Narenthiran, Sean Ficek, Aleksander Samadi, Mehrzad Huang, Jocelyn Jain, Siddhartha Gitman, Igor Moshkov, Ivan Du, Wei Toshniwal, Shubham Armstrong, George Kisacanin, Branislav Novikov, Matvei Gitman, Daria Bakhturina, Evelina Varshney, Prasoon Narsimhan, Makesh Scowcroft, Jane Polak Kamalu, John Su, Dan Kong, Kezhi Kliegl, Markus Mahabadi, Rabeeh Karimi Lin, Ying Satheesh, Sanjeev Parmar, Jupinder Gundecha, Pritam Norick, Brandon Jennings, Joseph Prabhumoye, Shrimai Akter, Syeda Nahida Patwary, Mostofa Khattar, Abhinav Narayanan, Deepak Waleffe, Roger Zhang, Jimmy Su, Bor-Yiing Huang, Guyue Kong, Terry Chadha, Parth Jain, Sahil Harvey, Christine Segal, Elad Huang, Jining Kashirsky, Sergey McQueen, Robert Putterman, Izzy Lam, George Venkatesan, Arun Wu, Sherry Nguyen, Vinh Kilaru, Manoj Wang, Andrew Warno, Anna Somasamudramath, Abhilash Bhaskar, Sandip Dong, Maka Assaf, Nave Mor, Shahar Argov, Omer Ullman Junkin, Scot Romanenko, Oleksandr Larroy, Pedro Katariya, Monika Rovinelli, Marco Balas, Viji Edelman, Nicholas Bhiwandiwalla, Anahita Subramaniam, Muthu Ithape, Smita Ramamoorthy, Karthik Wu, Yuting Velury, Suguna Varshini Almog, Omri Daw, Joyjit Fridman, Denys Galinkin, Erick Evans, Michael Ghosh, Shaona Luna, Katherine Derczynski, Leon Pope, Nikki Long, Eileen Schneider, Seth Siman, Guillermo Grzegorzek, Tomasz Ribalta, Pablo Katariya, Monika Alexiuk, Chris Conway, Joey Saar, Trisha Guan, Ann Pawelec, Krzysztof Prayaga, Shyamala Kuchaiev, Oleksii Ginsburg, Boris Olabiyi, Oluwatobi Briski, Kari Cohen, Jonathan Catanzaro, Bryan Alben, Jonah Geifman, Yonatan Chung, Eric Computation and Language Artificial Intelligence Machine Learning We introduce the Llama-Nemotron series of models, an open family of heterogeneous reasoning models that deliver exceptional reasoning capabilities, inference efficiency, and an open license for enterprise use. The family comes in three sizes -- Nano (8B), Super (49B), and Ultra (253B) -- and performs competitively with state-of-the-art reasoning models such as DeepSeek-R1 while offering superior inference throughput and memory efficiency. In this report, we discuss the training procedure for these models, which entails using neural architecture search from Llama 3 models for accelerated inference, knowledge distillation, and continued pretraining, followed by a reasoning-focused post-training stage consisting of two main parts: supervised fine-tuning and large scale reinforcement learning. Llama-Nemotron models are the first open-source models to support a dynamic reasoning toggle, allowing users to switch between standard chat and reasoning modes during inference. To further support open research and facilitate model development, we provide the following resources: 1. We release the Llama-Nemotron reasoning models -- LN-Nano, LN-Super, and LN-Ultra -- under the commercially permissive NVIDIA Open Model License Agreement. 2. We release the complete post-training dataset: Llama-Nemotron-Post-Training-Dataset. 3. We also release our training codebases: NeMo, NeMo-Aligner, and Megatron-LM.
title	Llama-Nemotron: Efficient Reasoning Models
topic	Computation and Language Artificial Intelligence Machine Learning
url	https://arxiv.org/abs/2505.00949