Amazon Web Services (AWS) on the 4th (local time) unveiled new features at "AWS re:Invent 2025" that lower the barrier to customizing artificial intelligence (AI) models.
By adding reinforcement learning-based customization to Amazon Bedrock and Amazon SageMaker AI, even developers without Machine Learning expertise can easily perform advanced model tuning.
The main challenge for AI corporations is efficiency. Running large-scale models brings high expense and slow response speeds. To reduce this inefficiency, AWS proposed a "custom small-model strategy" that helps use small models for frequently repeated simple tasks.
To that end, AWS introduced "RFT (Reinforcement Fine Tuning)" in Amazon Bedrock. RFT automates the complex pipelines of conventional Reinforcement Learning and human feedback, so once developers specify a dataset and a reward function, the model is fine-tuned automatically. According to AWS, RFT delivers an average 66% accuracy improvement over base models and will be applied starting with the Amazon Nova 2 Lite model. Salesforce, Inc. said the performance of its own AI agent improved by up to 73% after adopting RFT.
Meanwhile, SageMaker AI gained a "serverless reinforcement learning-based model customization" feature. It automates model tuning that used to take months so it can be completed in a few days. Developers can choose between an "agentic" mode, where an agent guides the entire process, and a "self-guided" mode for direct control. Both approaches support advanced techniques such as reinforcement learning from AI feedback (RLAF), verifiable reward learning, supervised fine-tuning, and direct preference optimization (DPO). This feature is compatible not only with Amazon Nova but also with open-weight models such as Llama, Qwen, and DeepSeek.
AWS said, "Now organizations of all sizes can build custom AI more easily and quickly," and emphasized, "Our goal is to raise AI's efficiency and accessibility at the same time."