At re:Invent 2025, AWS announced major updates to its Amazon Bedrock and Amazon SageMaker AI platforms, focusing on simplifying the advanced model customization process.
This update will address the "efficiency" and "cost" challenges faced by enterprises when deploying AI agents. By introducing "Reinforcement Fine Tuning" (RFT) and "Serverless Model Customization" features, developers can easily train more accurate, lightweight, and cost-effective customized models without having a deep background in machine learning, thereby replacing the use of expensive large general-purpose models.
Amazon Bedrock RFT: Enables models to learn from feedback, improving accuracy by 66%.
The Reinforcement Fine-Tuning (RFT) mechanism, introduced on Amazon Bedrock, simplifies the previously extremely complex reinforcement learning process. Developers simply select a base model (initially supporting Amazon Nova 2 Lite), import historical records or datasets, and then set a reward function (such as AI scoring or rules), and the system will automatically handle the end-to-end fine-tuning.
AWS points out that strengthening the fine-tuning mechanism can improve model accuracy by an average of 66%. Salesforce's test data further shows that, for specific business needs, strengthening the fine-tuning mechanism can improve model accuracy by up to 73%, while maintaining its high-quality execution and security.
SageMaker AI: Serverless customization reduces work time from months to days.
For developers requiring finer-grained control, SageMaker AI introduces a new serverless model customization feature. This feature offers two modes:
• Agentic Experience:The AI agent guides developers through the entire process from generating synthetic data to evaluation, requiring only a description of the requirements in natural language.
• Self-guided Approach:It provides advanced users with the flexibility to fine-tune parameters while relieving them of the burden of infrastructure management.
This feature supports a variety of models, including Amazon Nova, Llama, Qwen, DeepSeek, and GPT-OSS. Collinear AI, an AI technology company, has used this feature to shorten its model fine-tuning cycle from several weeks to several days.
SageMaker HyperPod: Checkpoint-free training, fault recovery in minutes
In addition, AWS has also launched the Checkpointless Training feature to meet the needs of large-scale model training.
In the past, if a failure occurred while thousands of accelerators were training, recovery from the checkpoint could take up to an hour. The new technology, however, can continuously save the model state, and once a failure occurs, the system will automatically resume training from a nearby healthy accelerator via peer-to-peer transmission, thereby significantly reducing the downtime to a few minutes and maintaining the utilization rate of expensive computing power clusters at 95%.
