Serverless Model Customization in Amazon SageMaker AI
Explore the models, techniques, and capabilities available with serverless model customization in Amazon SageMaker AI
Customizing foundation models has traditionally required navigating complex infrastructure decisions, from choosing the right instance types and configuring distributed training to managing GPU clusters and optimizing for cost. Serverless model customization in Amazon SageMaker AI removes this undifferentiated heavy lifting for fine-tuning and evaluation. You select a model, choose a customization technique, point to your data, and SageMaker AI automatically provisions the appropriate compute resources based on the model and data size. There's no training infrastructure to configure, manage, or scale. Training and validation metrics are logged to serverless MLflow, and when your job completes you can deploy to a SageMaker AI endpoint or Amazon Bedrock.
This post provides an overview of what's available today, and we'll keep it updated as we add support for new models and features.
Supported Techniques
Serverless model customization supports the following techniques:
Supervised Fine-Tuning (SFT)
SFT trains the model on curated input-output pairs so it learns to replicate desired patterns. This is the most straightforward technique and works well when you have high-quality labeled examples of the behavior you want.
Direct Preference Optimization (DPO)
DPO trains the model using pairs of preferred and rejected responses. Instead of learning from a single correct answer, the model learns which responses are better than others, making it effective for aligning model behavior with human preferences.
Reinforcement Learning with Verifiable Rewards (RLVR)
RLVR uses a reward function you define to score model-generated responses. For each prompt, the model generates multiple candidates, the reward function scores each one, and the model updates its policy to favor higher-scoring responses using Group Relative Policy Optimization (GRPO). This technique works well for tasks with objectively verifiable outcomes like tool calling, code generation, and structured output.
Reinforcement Learning from AI Feedback (RLAIF)
RLAIF uses an AI judge model to evaluate responses instead of a hand-written reward function. This is useful when correctness is harder to define programmatically, such as tone, helpfulness, or instruction following.
Supported Models
Provider | Models |
|---|---|
Amazon | Nova 2.0 Lite, Nova Pro, Nova Lite, Nova Micro |
Meta | Llama 3.3 70B Instruct, Llama 3.1 8B Instruct, Llama 3.2 3B Instruct, Llama 3.2 1B Instruct |
Qwen | Qwen3-32B, Qwen3-14B, Qwen3-8B, Qwen3-4B, Qwen3-1.7B, Qwen3-0.6B, Qwen2.5-72B-Instruct, Qwen2.5-32B-Instruct, Qwen2.5-14B-Instruct, Qwen2.5-7B-Instruct |
Deepseek AI | DeepSeek-R1-Distill-Llama-70B, DeepSeek-R1-Distill-Llama-8B, DeepSeek-R1-Distill-Qwen-32B, DeepSeek-R1-Distill-Qwen-14B, DeepSeek-R1-Distill-Qwen-7B, DeepSeek-R1-Distill-Qwen-1.5B |
OpenAI | gpt-oss-120b, gpt-oss-20b |
Last updated: March 2026
Getting Started
You can get started in Amazon SageMaker Studio in a few clicks:
- Choose Models in the left navigation pane
- Select a model and choose Customize with UI
- Select your technique, point to your training data, and configure hyperparameters
- Launch your customization job
Serverless model customization handles training and evaluation infrastructure so you can focus on your data and your model.
For training and evaluation data format requirements, see dataset formats in the documentation . For pricing details, see SageMaker AI pricing .
Evaluation
After training, evaluate your fine-tuned model against the base model directly from the model details page. Serverless model customization supports several evaluation approaches:
Standard NLP Metrics
Built-in metrics like ROUGE, F1, and exact match measure text quality and accuracy against your held-out evaluation set. These provide a quick baseline comparison between your fine-tuned and base models.
Custom Metrics
Define your own evaluation metrics tailored to your use case. For example, a tool calling reward function that verifies whether the model selected the correct function with the right parameters, or a code evaluation metric that checks whether generated code passes test cases.
LLM-as-a-Judge
Use a foundation model to evaluate your fine-tuned model's responses. This is useful for subjective qualities like helpfulness, coherence, and instruction following that are difficult to capture with programmatic metrics.
Standard Benchmarks
Evaluate against established benchmarks to measure general model capabilities and compare performance across models.
You can combine multiple evaluation approaches in a single evaluation job to get a comprehensive view of your model's performance.
When you're satisfied with the results, choose Deploy and select your deployment target: either a SageMaker AI endpoint or Amazon Bedrock. You can also download the model weights from S3 for self-managed deployment.
Conclusion
Serverless model customization in Amazon SageMaker AI lets you fine-tune and evaluate FMs using SFT, DPO, RLVR, and RLAIF, without managing training infrastructure.
For a hands-on walkthrough, see Serverless Model Customization with RLVR in SageMaker AI . To learn more, see the SageMaker AI model customization documentation .
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article