Nemotron Elastic: Towards Efficient Many-in-One Reasoning LLMs

Ali Taghibakhshi*, Sharath Turuvekere Sreenivas*, Saurav Muralidharan*, Ruisi Cai†, Marcin Chochowski, Ameya Sunil Mahabaleshwarkar, Yoshi Suhara, Oluwatobi Olabiyi, Daniel Korzekwa, Mostofa Patwary, Mohammad Shoeybi, Jan Kautz, Bryan Catanzaro, Ashwath Aithal, Nima Tajbakhsh, Pavlo Molchanov · 2025

Read Nemotron Elastic: Towards Efficient Many-in-One Reasoning LLMs by Ali Taghibakhshi*, Sharath Turuvekere Sreenivas*, Saurav Muralidharan*, Ruisi Cai†, Marcin Chochowski, Ameya Sunil Mahabaleshwarkar, Yoshi Suhara, Oluwatobi Olabiyi, Daniel Korzekwa, Mostofa Patwary, Mohammad Shoeybi, Jan Kautz, Bryan Catanzaro, Ashwath Aithal, Nima Tajbakhsh, Pavlo Molchanov (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-16664.txt.

Description

Ali Taghibakhshi and fifteen co-authors present Nemotron Elastic, a framework for training one reasoning language model that contains several smaller models inside it. Training a separate model for every deployment size is prohibitively expensive, and standard pruning or distillation still burns hundreds of billions of tokens per compressed model.

Nested submodels share the parent's weights and can be extracted at deployment with no extra training, selected by an end-to-end trained router and a two-stage curriculum built for reasoning models. Applied to a twelve-billion-parameter parent, the method produced nine- and six-billion-parameter models using only about 110 billion training tokens. That is a cost cut of more than three hundredfold versus training a family from scratch. Each nested model matched or beat the state of the art at its size.

Catalog record

CREATORAli Taghibakhshi*, Sharath Turuvekere Sreenivas*, Saurav Muralidharan*, Ruisi Cai†, Marcin Chochowski, Ameya Sunil Mahabaleshwarkar, Yoshi Suhara, Oluwatobi Olabiyi, Daniel Korzekwa, Mostofa Patwary, Mohammad Shoeybi, Jan Kautz, Bryan Catanzaro, Ashwath Aithal, Nima Tajbakhsh, Pavlo Molchanov
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE63 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-16664

Browse the full library · About the Tower of Babel