AI and ModelsConfirmedAWS SageMaker AI has launched new training session APIs, giving developers unprecedented control over serverless AI model customization, including custom training loops for advanced techniques like reinforcement learning, without managing GPU infrastructure.
On October 22, 2026, Amazon SageMaker AI announced the general availability of training session APIs for its serverless model customization offering. These APIs are available in US East (N. Virginia), US West (Oregon), and Asia Pacific (Mumbai) regions.
The new training session APIs provide customers with direct control over their post-training algorithms without requiring them to provision or manage GPU infrastructure. This expands upon existing SageMaker AI capabilities, which previously offered managed techniques like Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and other reinforcement learning (RL) methods where SageMaker AI handled the training execution based on customer-selected techniques and models.
With the introduction of training session APIs, developers can now write custom training loops. This enables them to reproduce existing techniques like SFT or DPO on their own terms, or to implement custom loss functions, new RL reward signals, data curricula, or algorithms derived from recent research. Specific operations, such as `forward_backward()`, `optim_step()`, and `sample()`, are exposed as API calls against a live, stateful session.
SageMaker AI manages the underlying infrastructure and operational aspects for these sessions, including GPU allocation, loading of models and LoRA (Low-Rank Adaptation) weights, distributed execution, fault recovery, checkpoint storage, and security. Developers interact with these operations directly from their Python code to drive custom training objectives, including reinforcement learning and supervised fine-tuning, with built-in checkpointing and resume functionality.
The billing model for these training sessions is per token processed. Local computation, such as reward scoring in RL, is not billed as GPU time. To assist developers, SageMaker AI has released the SageMaker training sessions cookbook, a collection of ready-to-run SFT and RL recipes that serve as blueprints for custom training loops. Getting started involves installing the SageMaker Python SDK, cloning the cookbook, or utilizing a coding agent within SageMaker Studio.
The introduction of training session APIs for Amazon SageMaker AI represents a significant shift in the balance between abstraction and control for AI model customization. Previously, SageMaker AI offered a high-level, managed approach where developers selected a technique and a model, and the service handled the training execution. While convenient for standard use cases, this abstraction limited the ability to implement highly specialized or experimental training methodologies. The new APIs bridge this gap, offering granular control over the training loop while retaining the benefits of a serverless infrastructure model.
This development directly addresses a common challenge in advanced AI engineering: the need for custom training logic that deviates from predefined frameworks or managed services. Techniques like novel reinforcement learning algorithms, complex data curricula, or research-driven loss functions often require direct manipulation of the forward and backward passes, optimizer steps, and data sampling. By exposing operations such as `forward_backward()`, `optim_step()`, and `sample()` as API calls, SageMaker AI provides the necessary primitives for developers to construct these custom loops. This allows AI engineers to focus on algorithmic innovation rather than the operational complexities of distributed GPU training.
The serverless nature of these training sessions is a key engineering consequence. SageMaker AI explicitly handles GPU allocation, model and LoRA loading, distributed execution, fault recovery, checkpoint storage, and security. This abstraction removes substantial operational overhead for developers, who no longer need to manage EC2 instances, container orchestration, GPU drivers, or distributed training frameworks. For teams that previously maintained custom GPU clusters or complex MLOps pipelines for fine-tuning, this can significantly reduce infrastructure management costs and accelerate iteration cycles. The ability to interact with a live, stateful session via Python code integrates well with existing data science workflows and tooling.
The billing model, based on tokens processed rather than GPU time, introduces a distinct cost optimization vector. For training procedures that involve substantial local computation (e.g., complex reward function calculations in reinforcement learning, extensive data augmentation, or pre-processing steps that run on CPU), this model can lead to more efficient resource utilization and lower costs. Developers are incentivized to offload as much non-GPU-intensive computation as possible to their local environment or to the CPU component of the serverless session, paying only for the core GPU operations that process tokens. This contrasts with traditional GPU-hour billing, where all computation within the allocated GPU instance contributes to cost, regardless of whether the GPU is actively processing tensors.
The SageMaker training sessions cookbook, providing ready-to-run SFT and RL recipes, serves as a critical enabler for adoption. These blueprints offer practical examples of how to construct custom training loops using the new APIs, lowering the barrier to entry for developers unfamiliar with this level of control within a managed service. For AI researchers and engineers pushing the boundaries of model customization, this capability facilitates rapid experimentation and the implementation of cutting-edge techniques without the prohibitive setup and maintenance costs associated with self-managed infrastructure. The focus shifts from infrastructure engineering to algorithm engineering, potentially accelerating the development of more specialized and performant AI models tailored to specific domain requirements. While the APIs offer granular control over the training loop, the underlying infrastructure remains managed by AWS, meaning developers do not have direct control over aspects like specific hardware configurations or operating system environments; their control is focused on the algorithmic steps within the provided execution environment.