NVIDIA paper: Nemotron-H, a family of hybrid Mamba-Transformer models

cwolferesearch · x · 2026-08-19

NVIDIA released the Nemotron-H paper, detailing a hybrid Mamba-Transformer architecture for accuracy and efficiency. The accompanying discussion highlights midtraining/CPT techniques: emphasizing high-quality domain-specific data (math, code, science) while retaining general data to avoid catastrophic forgetting.

Original post →

More from Models

Models channel →