NVIDIA Launches NeMo AutoModel to Accelerate MoE AI Training

NVIDIA has launched NeMo AutoModel, a software update designed to accelerate the training of mixture-of-experts AI models. The tool expands the company’s NeMo framework and integrates directly with Hugging Face’s Transformers v5, which introduced native support for the MoE architecture. By leveraging expert parallelism, DeepEP communication kernels, and TransformerEngine optimizations, the software targets developers working within the open-source Transformers ecosystem, allowing them to plug in performance upgrades with minimal code changes. NVIDIA stated the update increases training throughput by 3.4 to 3.7 times compared to previous configurations.

Mixture-of-experts models have become a standard approach for scaling large language models while managing computational costs. These architectures route data through specialized subsets of parameters rather than processing every input with a model’s full capacity. The NeMo AutoModel release follows Hugging Face’s integration of MoE support in late June, positioning NVIDIA’s optimization as a direct response to the growing industry demand for faster, more efficient AI training pipelines.