Model Customization with Amazon SageMaker AI
Customize models with your data using the broadest set of techniques. Fully serverless, on the infrastructure you trust.
Why SageMaker AI for model customization
Amazon SageMaker AI enables AI developers to customize 20+ popular open-weight models and Amazon Nova using supervised fine-tuning (SFT), direct preference optimization (DPO), reinforcement learning with verifiable rewards (RLVR), and reinforcement learning from AI feedback (RLAIF).
The entire workflow is fully serverless — SageMaker handles compute provisioning, scaling, and optimization.
Deploy custom models to Amazon Bedrock for serverless inference or SageMaker endpoints for managed inference. See the documentation page for the latest list of supported models and techniques
Benefits
Fast track model customization with maximum accuracy
Complete the end-to-end workflow from data preparation to deployment in days, not months. Get started with the guided UI or an AI agent-guided workflow, then automate with the SageMaker Python SDK. Fully serverless — no infrastructure to provision, no capacity to manage.
A broad selection of customization techniques across 20+ open-weight models, all through a serverless experience. Open-weight models mean you own the weights — deploy your custom model to Amazon Bedrock for serverless inference, SageMaker endpoints for managed inference, or export to your own infrastructure. Model customization skills are open-source on GitHub, so teams can inspect, fork, and adapt them to fit their workflows.
Training metrics and evaluation results are tracked and logged directly in MLflow on SageMaker AI, so you can monitor every experiment and compare model performance with full visibility.
Model customization made simple
Comprehensive capabilities to customize models across the end-to-end workflow
Data preparation
Using model customization skills, your coding agent generates code to format your data for your selected model and customization technique, validates data quality, and identifies gaps. This reduces weeks of manual data curation to hours. For broader data preparation needs, SageMaker AI offers Data Wrangler for visual data transformation with 300+ built-in transforms, Ground Truth for data labeling at scale, and Processing jobs for custom data processing workflows. Bring your own data or generate synthetic data (in preview) when real-world data is limited, sensitive, or difficult to obtain. SageMaker AI can generate task-specific synthetic training examples that complement your existing datasets, helping you overcome data scarcity without compromising on model quality.
Advanced customization techniques
SageMaker AI supports the latest model customization techniques including supervised fine-tuning (SFT), direct preference optimization (DPO), and reinforcement learning from AI feedback (RLAIF) and verifiable rewards (RLVR).
SageMaker AI supports supervised fine-tuning (SFT) or direct preference optimization (DPO) when you need to change model behavior, and advance to reinforcement learning (RLVR, RLAIF) when you need to optimize for specific reward signals. For continued pre-training, leverage SageMaker Training Jobs or SageMaker HyperPod to scale across thousands of accelerators.
End-to-end serverless model customization
SageMaker AI automatically selects and provisions the appropriate compute resources based on the model and data size—all without requiring you to select and manage instances.
Inference
Once you have achieved your desired accuracy and performance goals, you can deploy models to production in a few clicks to either SageMaker AI inference endpoints or Amazon Bedrock for serverless inference
LLMOps
You can automatically log all critical experiment metrics all without provisioning a tracking server or modifying code. Integration with MLflow also provides rich visualizations and an ingress into the MLflow user interface for further analysis.
Easy-to-use interface
Use the guided UI in SageMaker Studio to select a model, choose a technique, configure your dataset, launch serverless training, monitor live metrics, and deploy — all in a few clicks. No agent or code required.
Agent-guided development
Describe your use case and an AI coding agent guides you through data transformation, technique selection, hyperparameter configuration, evaluation, and deployment. Purpose-built agent skills bring specialized expertise in fine-tuning techniques, model selection, SageMaker AI APIs, and evaluation methodologies to your AI coding agent, helping you develop faster with confidence. Agent skills are pre-built, customizable instruction sets that fit your existing workflows and governance standards. Get started in SageMaker Studio JupyterLab with Kiro and agent skills preinstalled or use agent skills with your preferred IDE or coding agent, including Kiro IDE, Cursor, Claude Code, and VS Code.
SageMaker Python SDK
Programmatically configure and launch customization jobs, define evaluation criteria, and deploy models. Full control over every parameter for developers who prefer code-first workflows.
Customers
Why our customers choose Amazon SageMaker AI for model customization
Collinear AI
"At Collinear, we build curated datasets and simulation environments for frontier AI labs and Fortune 500 enterprises to improve their models. Fine-tuning AI models is critical to creating high-fidelity simulations, and it used to require stitching together different systems for training, evaluation, and deployment. Now with Amazon SageMaker AI's new serverless model customization capability, we have a unified way that empowers us to cut our experimentation cycles from weeks to days. This end-to-end serverless tooling helps us focus on what matters: building better training data and simulations for our customers, not maintaining infrastructure or juggling disparate platforms."
Soumyadeep Bakshi, Co-founder, Collinear AI
Oleum
"At Oleum, we're building AI tools that help organizations understand and trust their data. Amazon SageMaker AI's new agentic AI experience is exactly the kind of tooling we need. It acts as a thought partner, not just a task executor — recommending techniques, catching mismatches in our data, and letting us build any workflow we want rather than forcing us through a rigid process. The fact that these skills plug directly into our existing agentic development environment means we can experiment with fine-tuning approaches without context-switching between platforms. And with the new chat experience built directly into JupyterLab in SageMaker Studio, we can go from conversation to runnable notebook to training job all in one place. It's flexible, customizable, and built for how modern ML teams actually work.”
Oleum — Alejandro Ballesteros, CTO
Wink
"At Wink, we're building AI-powered digital twins that capture the nuances of real human personalities, helping users screen for genuine connections before meeting in person. Our development workflow has shifted entirely to agentic, prompt-driven experiences — we move fast and ship fast. Amazon SageMaker AI's new skills-based approach to model customization fits perfectly into that workflow. Instead of wrestling with infrastructure or rigid interfaces, our team can fine-tune personality models through natural language right inside the tools we already use. For a startup where speed and cost efficiency are everything, this is a game-changer — it lets us focus on building better experiences for our users, not on managing ML pipelines.”
Wink — Ethan Fan, CTO
Robin AI
"At Robin, we're redefining the role of legal in modern business and using AI to drive better decisions, faster actions, and sustainable growth. To provide our clients with better decision making, it is crucial that our AI models match how lawyers write contracts—from the specific format, tone, and preferences of individual lawyers. Previously, customizing models with proprietary data was a cumbersome process that was prone to errors. Now with the new serverless model customization capability in Amazon SageMaker AI, we can rapidly experiment with advanced techniques like reinforcement learning with verifiable rewards in just days. In addition, we're excited to try the AI agent-guided workflow so we can compare and verify our assumptions to help lawyers around the globe make better decisions faster.“
Diana Mincu - Director of Research, Robin AI
Did you find what you were looking for today?
Let us know so we can improve the quality of the content on our pages