Breaking

NVIDIA Unveils NeMo AutoModel, Claiming Up to 3.7x Faster MoE Training

June 24, 2026 at 12:03 EDT

Open Source
Infra & Chips
Foundation Models

On June 24, 2026, NVIDIA introduced "NeMo AutoModel," an open-source training library built on top of Hugging Face's Transformers v5, claiming it boosts training throughput for Mixture-of-Experts (MoE) models by up to 3.4–3.7x.

NVIDIA NeMo · Open-Source Training

NeMo AutoModel: up to 3.7× faster MoE training, with Day-0 Hugging Face support

An open-source library on Hugging Face Transformers v5 that loads any Hub model instantly — no checkpoint conversion — and layers NVIDIA optimizations (Expert Parallelism, DeepEP, TransformerEngine) in a few lines of code.

3.4–3.7×

Training throughput boost for popular MoE models

Day-0

Any Hugging Face Hub model trains instantly — zero conversion

1 → 1000s

Scales from a single GPU to thousands with one config

Benchmark throughput — tokens/sec/GPU

PyTorch backend, NVIDIA GPUs. Column height proportional to per-GPU token rate.

13,058

GPT-OSS 20B

8 GPUs · 279 TFLOPs

12,040

Qwen3 MoE 30B

8 GPUs · 277 TFLOPs

1,002

DeepSeek V3 671B

256 GPUs · 250 TFLOPs

The hybrid positioning

HF + FSDP

Easy, but slower

Megatron-Core

Fast, but conversion cost

→

NeMo AutoModel

Day-0 HF + NVIDIA speed

What developers value

Fine-tune HF models with no conversion
PyTorch-native: drops into TRL & lm-eval-harness
Recipes combine FSDP2 + EP + PP + CP
New models added almost weekly

Caveats to weigh

Large MoE may need Expert Parallelism tuning
Small single-node experiments can incur overhead
Best suited to rapid experimentation at scale

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.