Software

Amazon SageMaker Enables Serverless Customization of NVIDIA Nemotron 3 Models

Amazon SageMaker now allows serverless fine-tuning of NVIDIA Nemotron 3 models, including the 30B and 120B parameter variants, to tailor AI for enterprise use.

A man with a prosthetic hand gently touches a woman's face in a cyberpunk setting.

Photo: Yaroslav Shuraev / Pexels

Amazon SageMaker AI has introduced serverless model customization for NVIDIA Nemotron 3 models, enabling businesses to fine-tune these open-weight large language models for specific domains and workflows. This approach allows organizations to adapt the models to their unique needs without managing infrastructure. The new feature supports models like Nemotron 3 Nano (30B total parameters, 3B active) and Nemotron 3 Super (120B total parameters, 12B active), offering flexibility for enterprises seeking specialized AI solutions. The customization process uses techniques such as supervised fine-tuning, reinforcement learning with verifiable rewards, and reinforcement learning from AI feedback to align models with business requirements. This marks a significant step in making AI more adaptable for enterprise applications, as it allows organizations to create proprietary intellectual property tailored to their operations.

The NVIDIA Nemotron 3 models are built on a hybrid Mamba-Transformer Mixture-of-Experts (MoE) architecture, which enables efficient processing of long-context sequences. This design activates only a fraction of total parameters per forward pass, such as 12B of 120B in the Super variant, reducing compute costs while maintaining high throughput and accuracy. The models are trained using multi-environment reinforcement learning through NeMo Gym, which helps align them with real-world, multi-step tasks across domains like coding, reasoning, and long-context analysis. These features make the Nemotron 3 models particularly effective for complex tasks that require both reasoning and efficiency.

The serverless customization feature in Amazon SageMaker AI removes the need for users to provision GPU clusters or manage distributed training frameworks. Instead, SageMaker AI handles infrastructure and training orchestration, allowing users to focus on their data and business use cases. This approach ensures that organizations can pay only for the resources they use, improving cost efficiency. The customization options include techniques such as supervised fine-tuning, which uses labeled input-output pairs to teach the model new behaviors, and reinforcement learning from AI feedback, which uses an AI model to guide optimization without human-labeled data.

Source: awsml

Key points

Amazon SageMaker AI now allows serverless fine-tuning of NVIDIA Nemotron 3 models, including the 30B and 120B parameter variants.
Nemotron 3 Nano (30B total parameters, 3B active) and Nemotron 3 Super (120B total parameters, 12B active) are supported for serverless model customization.
The models use a hybrid Mamba-Transformer Mixture-of-Experts (MoE) architecture to process long-context sequences efficiently.
Nemotron 3 models are trained using multi-environment reinforcement learning through NeMo Gym to align with real-world tasks.
SageMaker AI serverless customization removes the need for users to manage GPU clusters or distributed training frameworks.
Supervised fine-tuning (SFT) uses labeled input-output pairs to teach the model new behaviors.
Reinforcement learning from AI feedback (RLAIF) uses an AI model to guide optimization without human-labeled data.

Source: AWS Machine Learning Read the original →

WRITTEN BY

Theo Almeida

AI Software & Developer Tools

Theo covers AI software, developer tools, frameworks, and the platforms builders use every day.