- What: AWS expanded Reinforcement Fine-Tuning (RFT) support to open-weight models via OpenAI-compatible APIs.
- When: Initially launched for Amazon Nova in December 2025; extended to open-weight models in February 2026.
- Models Supported: OpenAI GPT OSS 20B, Qwen 3 32B, Amazon Nova, and Llama.
- Technical Edge: Uses the GRPO algorithm and AWS Lambda-based reward functions to automate model improvement without massive labeled datasets.
Amazon Web Services (AWS) has significantly lowered the technical hurdle for enterprise AI customization by introducing Reinforcement Fine-Tuning (RFT) for open-weight models on Amazon Bedrock. By utilizing OpenAI-compatible APIs, developers can now fine-tune powerful models like OpenAI’s GPT OSS 20B and Qwen 3 32B using an iterative feedback loop rather than traditional, labor-intensive supervised datasets.
This announcement, following the initial rollout of RFT for Amazon Nova models in late 2025, represents a fundamental shift in how large language models (LLMs) are optimized for specialized tasks. Instead of requiring thousands of static input-output pairs, RFT allows models to "learn by doing," receiving numerical rewards for successful responses in complex domains such as mathematical reasoning and code generation.
Beyond Supervised Learning: The RFT Advantage
Traditional Supervised Fine-Tuning (SFT) has long been the industry standard, but it suffers from a significant bottleneck: the need for high-quality, human-labeled data. RFT breaks this cycle by treating the LLM as an "agent" that explores various responses to a single prompt.
According to AWS technical documentation, RFT functions similarly to training a chess player. Rather than showing the model every possible move in every situation, the system allows the model to play and provides feedback on which moves lead to a "winning" position. In the context of Bedrock, the model generates multiple candidate responses, which are then scored by a reward function.
This "online learning" capability allows the model to adapt in real-time. As the model improves, it encounters new, more challenging scenarios, leading to a compounding effect on performance. For verifiable tasks like math, where correctness can be checked automatically, this process eliminates the need for expensive manual labeling entirely.
Seamless Integration via OpenAI-Compatible APIs
To drive adoption, AWS has engineered the RFT workflow to be compatible with the OpenAI Chat Completions format. This allows developers to transition existing workloads to Amazon Bedrock without rewriting substantial portions of their codebase.
The technical workflow involves several key components:
- The Actor Model: The foundation model being customized (e.g., Qwen 3 32B or GPT OSS 20B).
- Reward Functions: Implemented as AWS Lambda functions, these evaluate model responses and return numerical scores.
- GRPO Algorithm: Bedrock utilizes Group Relative Policy Optimization (GRPO), a state-of-the-art reinforcement learning algorithm designed for policy optimization.
- Automated Infrastructure: The platform handles batching, parallelization, and resource allocation, with built-in convergence detection to stop training automatically once optimal performance is reached.
Developers can launch these jobs by specifying the base model, dataset, and reward function through the Bedrock console or API. Once fine-tuning is complete, the resulting model is immediately available for on-demand inference through the same OpenAI-compatible Chat Completions and Responses APIs.
Impact on Developers and the AI Industry
For developers, this release means that sophisticated reinforcement learning—previously reserved for labs with massive compute resources—is now accessible as a managed service. By supporting open-weight models like Qwen 3 and GPT OSS 20B, AWS is positioning Bedrock as a neutral, high-performance ground for model customization.
"RFT in Amazon Bedrock automates the end-to-end customization workflow," the company stated in its technical walkthrough. This automation is critical for enterprises that need to improve model accuracy in niche areas like the GSM8K math dataset or proprietary unit testing for software development.
The impact section can be summarized by one core shift: Efficiency translates to power. By reducing the reliance on pre-collected examples, companies can iterate faster and deploy models that are significantly more accurate in multi-turn conversations and reasoning-heavy tasks.
What’s Next for Amazon Bedrock
The expansion to open-weight models in February 2026 suggests a roadmap where Amazon Bedrock continues to embrace the open-source and cross-platform AI community. As reward functions become more sophisticated—shifting from simple "LLM-as-Judge" setups to complex, verifiable unit tests—the gap between general-purpose models and domain-specific experts is expected to widen.
AWS has already integrated Amazon CloudWatch metrics into the RFT process, providing developers with real-time visibility into training progress. Future updates are likely to include broader support for additional open-weight architectures and even more granular control over hyperparameters within the OpenAI-compatible framework.