SageMaker AI now supports serverless fine-tuning for Nvidia Nemotron models
Fine-tune a 30B NVIDIA model on your own data in days—no GPU cluster management, no upfront cost, just results.
View original announcement →Visual Summary
graph TD
A{{Serverless Fine-Tuning Nemotron}}:::announced
B((Data Scientist)):::external
C(SageMaker Studio):::compute
D(SageMaker Python SDK V3):::compute
E([Supervised Fine-Tuning]):::feature
F([Reinforcement Fine-Tuning]):::feature
G(GPU Instances P5/P4/G5):::compute
H(Training Datasets):::storage
I([Logged Model Output]):::feature
B ==>|"launches job"| C
B -->|"programmatic access"| D
C ==>|"configures"| A
D ==>|"configures"| A
H -->|"proprietary data"| A
A -->|"provisions"| G
A -->|"applies"| E
A -->|"applies"| F
A -->|"produces"| I
G -.->|"auto cleanup"| A
classDef announced fill:#ff9900,stroke:#ec7211,color:#fff,font-weight:bold
classDef compute fill:#e3f2fd,stroke:#1565c0,color:#1565c0
classDef storage fill:#e8f5e9,stroke:#2e7d32,color:#2e7d32
classDef feature fill:#fff3e0,stroke:#e65100,color:#e65100
classDef external fill:#f5f5f5,stroke:#616161,color:#616161
What's New
Amazon SageMaker AI has added serverless model customization support for the NVIDIA Nemotron 3 Nano model, a 30-billion-parameter open-weight model, enabling both supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT). This capability allows organizations to adapt the Nemotron 3 Nano foundation model to domain-specific tasks, custom tones, or new workflows using proprietary data—without managing any underlying infrastructure. The feature is accessible via Amazon SageMaker Studio's guided UI or programmatically through the SageMaker Python SDK V3.
How It Works
- Model Selection: Users select the NVIDIA Nemotron 3 Nano (30B parameters) as the base foundation model from the Models page in SageMaker Studio or via the SageMaker Python SDK V3.
- Customization Techniques: Two fine-tuning methods are supported—Supervised Fine-Tuning (SFT), which trains the model on labeled prompt-response pairs, and Reinforcement Fine-Tuning (RFT), which uses reward signals to align model behavior with desired outcomes.
- Serverless Infrastructure: SageMaker AI automatically provisions GPU instances (P5, P4de, P4d, or G5) based on model size and training requirements, with no manual cluster configuration required.
- Pre-Optimized Recipes: The platform applies pre-built training recipes that encode best practices for each customization technique, reducing the need for hyperparameter tuning expertise.
- Data Integration: Users supply datasets—collections of training examples such as prompt-response pairs or domain-specific text—which are used to adapt the model's knowledge, style, or behavior.
- Monitoring and Cleanup: Real-time metrics and logs are accessible through the UI during training, and compute resources are automatically released after job completion to control costs.
- Output: The result is a "logged model"—a customized version of Nemotron 3 Nano that retains general capabilities while incorporating domain-specific adaptations, ready for deployment on SageMaker AI.
Why It's Important
- Democratizes Large-Model Fine-Tuning: Fine-tuning a 30B-parameter model traditionally requires significant MLOps expertise and GPU cluster management; serverless abstraction removes these barriers for teams without deep infrastructure skills.
- Accelerates Time-to-Value: The managed workflow compresses what was previously a months-long customization process into days, enabling faster iteration from proof-of-concept to production.
- Reduces Operational Cost and Risk: Pay-per-use pricing means organizations are not billed for idle GPU capacity, and automatic resource cleanup eliminates the risk of runaway compute costs.
- Enables Proprietary Data Utilization: Organizations can embed domain-specific knowledge, terminology, and behavioral guidelines into a powerful open-weight model without exposing data to third-party training pipelines.
- Expands NVIDIA-AWS Ecosystem: Adding Nemotron 3 Nano customization deepens the NVIDIA-AWS partnership, giving enterprises access to a high-quality open-weight model with enterprise-grade fine-tuning tooling on a single platform.
- Supports Compliance-Sensitive Workloads: Because customization runs within the customer's AWS environment, data governance and residency requirements can be maintained throughout the training process.
How It's Different
- Serverless vs. Self-Managed Training: Unlike traditional SageMaker Training Jobs where users must specify instance types, counts, and manage job lifecycle, serverless customization handles all provisioning and orchestration automatically.
- Open-Weight Model vs. Proprietary APIs: Nemotron 3 Nano is an open-weight model, giving users full ownership of the fine-tuned artifact and avoiding vendor lock-in associated with fine-tuning closed proprietary models via API.
- SFT + RFT vs. SFT-Only Offerings: Many managed fine-tuning services support only supervised fine-tuning; this feature also includes reinforcement fine-tuning (RFT), enabling more sophisticated alignment and reasoning improvements.
- SDK V3 Exclusive: The programmatic customization API is exclusive to SageMaker Python SDK V3, which was purpose-built for foundation model workflows, offering a more streamlined developer experience than legacy SDK patterns.
- Integrated Evaluation Assets: Unlike standalone fine-tuning pipelines, SageMaker model customization includes built-in evaluators and dataset management, reducing the need to stitch together separate evaluation tooling.
- Broader Technique Suite on the Platform: The underlying SageMaker customization platform also supports DPO, RLVR, RLAIF, and multi-turn RL for other models, positioning it as a more comprehensive customization hub than single-technique alternatives.
When to Prefer It
- Domain-Specific NLP Applications: When building applications in specialized verticals—such as legal research, medical documentation, or financial analysis—where a general-purpose model underperforms on domain terminology and reasoning patterns.
- Teams Without MLOps Infrastructure: When your organization lacks dedicated ML platform engineers to manage GPU clusters, distributed training frameworks, or job scheduling for large-model fine-tuning.
- Rapid Prototyping and Iteration: When you need to quickly validate whether fine-tuning improves task performance before committing to a full-scale, self-managed training infrastructure investment.
- Cost-Sensitive Projects: When training workloads are intermittent or experimental and paying for dedicated GPU capacity would be wasteful; serverless pay-per-use pricing is more economical in these scenarios.
- Alignment and Behavioral Customization: When the goal is not just knowledge injection but behavioral alignment—such as enforcing a specific tone, response format, or reasoning style—where RFT provides advantages over SFT alone.
- Regulated Industries Requiring Data Control: When compliance requirements mandate that proprietary training data remain within a controlled AWS environment rather than being sent to an external model provider's fine-tuning endpoint.
- Organizations Already Using SageMaker Studio: When your team's existing workflow is centered on SageMaker Studio, leveraging the integrated UI for customization avoids context-switching to separate tooling.
Availability
- GA Status: Generally available as of June 12, 2026.
- Supported Regions: US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo), and Europe (Ireland).
- Pricing Model: Serverless, pay-per-use—charges are incurred only for actual compute consumed during training; no cost for idle capacity.
- Supported Model: NVIDIA Nemotron 3 Nano (30B total parameters) is the supported model for this serverless customization feature at launch.
- Supported Techniques: Supervised Fine-Tuning (SFT) and Reinforcement Fine-Tuning (RFT) are available; the broader SageMaker platform also supports DPO, RLVR, and RLAIF for other models.
- SDK Requirement: Programmatic access requires SageMaker Python SDK V3; this feature is not available in earlier SDK versions.
- Access Methods: Available via the Models page in Amazon SageMaker Studio (guided UI) or the SageMaker Python SDK V3 for programmatic job submission.
- Infrastructure: Automatically provisions GPU instances from P5, P4de, P4d, and G5 families based on job requirements; instance selection is not user-configurable.
