Back to Research papers
Research paper index

Matryoshka Language Model Suites

Nathan Godey, Yoav Artzi

arXiv:2608.09703Published August 10, 20260 citations
  • cs.AI
  • cs.CL

Abstract

Training a language model suite classically requires training each model separately and serving them independently. We improve both training and inference efficiency by stacking sub-models of increasing size into a single nested architecture trained end-to-end. This Matryoshka training framework reduces the total parameter count of the suite, enables low-cost distillation from the largest to all smaller sub-models at every training step, and is well-suited for speculative decoding as the draft model is contained within the verifier. We validate our approach by training a Matryoshka suite comprising 500M, 1.5B, and 3B sub-models. Our suite is on par with independently trained baselines on benchmark performance and validation and out-of-domain perplexities, while using 36% less training compute and improving the throughput of speculative decoding by 14-26%. We also ablate key architectural choices, offering guidance for building strong Matryoshka LM suites.

Read the original paper

This page indexes public paper metadata. The manuscript remains with its original publisher and authors.