Today, Amazon SageMaker AI announces support for Seekable OCI (SOCI)-indexed container images for training and processing jobs, so your jobs start running your code sooner. SOCI-indexed images start Amazon SageMaker AI training containers in seconds, cutting GPU idle time during image pulls by up to 90%, with no impact on training performance. Before a training or processing job can run, SageMaker AI downloads its container image to the instance. For deep learning images, which routinely exceed 10 GB, that download adds minutes to the start of every job, and you pay for that time. Now, when your image includes a SOCI index, SageMaker AI lazy-loads the image, fetching only the files needed to start the container and streaming the rest in the background as your code runs. To use this feature, reference a SOCI-indexed image from a private Amazon ECR repository in your training or processing job. You can use a pre-indexed AWS Deep Learning Container by changing the image tag to its -soci variant, or add a SOCI index to your own image in Amazon ECR with open-source tools, without rebuilding the image. No changes to your job definition, entry point, or data channels are needed. SageMaker AI checks for an index automatically and images without an index download as before. If an index can't be used, the job falls back to a standard download and continues running. SOCI-indexed images for SageMaker AI training and processing jobs are available today in all AWS Regions where SageMaker AI is available, through the SageMaker AI APIs, CLIs, SDKs, and Console. To learn more, see the documentation for https://docs.aws.amazon.com/sagemaker/latest/dg/train-soci-indexed-images.html and https://docs.aws.amazon.com/sagemaker/latest/dg/processing-job-soci-indexed-images.html, or read the https://aws.amazon.com/blogs/machine-learning/announcing-soci-indexed-images-for-faster-sagemaker-ai-jobs/.

Today, Amazon SageMaker AI announces support for Seekable OCI (SOCI)-indexed container images for training and processing jobs, so your jobs start running your code sooner. SOCI-indexed images start Amazon SageMaker AI training containers in seconds, cutting GPU idle time during image pulls by up to 90%, with no impact on training performance. Before a training or processing job can run, SageMaker AI downloads its container image to the instance. For deep learning images, which routinely exceed 10 GB, that download adds minutes to the start of every job, and you pay for that time. Now, when your image includes a SOCI index, SageMaker AI lazy-loads the image, fetching only the files needed to start the container and streaming the rest in the background as your code runs. To use this feature, reference a SOCI-indexed image from a private Amazon ECR repository in your training or processing job. You can use a pre-indexed AWS Deep Learning Container by changing the image tag to its -soci variant, or add a SOCI index to your own image in Amazon ECR with open-source tools, without rebuilding the image. No changes to your job definition, entry point, or data channels are needed. SageMaker AI checks for an index automatically and images without an index download as before. If an index can't be used, the job falls back to a standard download and continues running. SOCI-indexed images for SageMaker AI training and processing jobs are available today in all AWS Regions where SageMaker AI is available, through the SageMaker AI APIs, CLIs, SDKs, and Console. To learn more, see the documentation for training jobs and processing jobs, or read the launch blog.

Amazon SageMaker AI now supports serverless model customization for the NVIDIA Nemotron 3.5 Lightning model using supervised fine-tuning (SFT), Direct Preference Optimization (DPO), and reinforcement fine-tuning (RFT). This is one of the latest open-weight models from NVIDIA and employs a hybrid Mixture-of-Experts architecture with 3B active parameters and 30B parameters in total. In addition to deploying this model on SageMaker AI, you can now adapt it to your specific domains and workflows. Model customization enables you to tailor foundation models with your proprietary data so a smaller, right-sized model can match frontier-model quality on your tasks, reducing cost and latency. You can use labeled data with SFT to improve accuracy on domain-specific tasks, preference data with DPO to align outputs with your organization's tone, or reward signals with RFT to enhance performance on new tasks. With serverless customization, SageMaker AI handles all infrastructure provisioning and training orchestration, so you can focus on your data and evaluation rather than cluster management, and only pay for what you use. Serverless model customization for NVIDIA Nemotron 3.5 Lightning on SageMaker AI is available in US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo), and Europe (Ireland). To get started, navigate to the Models page in Amazon SageMaker Studio to launch a customization job, or use the SageMaker Python SDK for programmatic access. To learn more, see the Amazon SageMaker AI model customization documentation.

Amazon SageMaker AI now supports serverless model customization for the NVIDIA Nemotron 3.5 Lightning model using supervised fine-tuning (SFT), Direct Preference Optimization (DPO), and reinforcement fine-tuning (RFT). This is one of the latest open-weight models from NVIDIA and employs a hybrid Mixture-of-Experts architecture with 3B active parameters and 30B parameters in total. In addition to deploying this model on SageMaker AI, you can now adapt it to your specific domains and workflows. Model customization enables you to tailor foundation models with your proprietary data so a smaller, right-sized model can match frontier-model quality on your tasks, reducing cost and latency. You can use labeled data with SFT to improve accuracy on domain-specific tasks, preference data with DPO to align outputs with your organization's tone, or reward signals with RFT to enhance performance on new tasks. With serverless customization, SageMaker AI handles all infrastructure provisioning and training orchestration, so you can focus on your data and evaluation rather than cluster management, and only pay for what you use. Serverless model customization for NVIDIA Nemotron 3.5 Lightning on SageMaker AI is available in US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo), and Europe (Ireland). To get started, navigate to the Models page in Amazon SageMaker Studio to launch a customization job, or use the https://sagemaker.readthedocs.io/en/stable/model_customization/index.html for programmatic access. To learn more, see the Amazon SageMaker AI https://docs.aws.amazon.com/sagemaker/latest/dg/customizing-models.html. 

Today, Amazon SageMaker AI announces instance preference lists for training and processing jobs, making it easier and faster to find compute capacity for your workloads. Many AI training, fine-tuning, and data processing workloads run comparably well on any of several instance types or sizes. However, before now, you had to name only one instance type at the time of job submission and wait for SageMaker to find that specific instance for your job. For high-demand GPUs during peak periods, where wait times can be unpredictable, customers sometimes had to build complex retry logic or concurrently submit multiple jobs with different instance types to find the first available option. Now you can simply provide a prioritized list of the instance types your workload accepts, and SageMaker automatically runs your job on the first available configuration from your preferences. With this solution, your training or processing job will likely start sooner. To use this feature, you specify your instance type and count preferences in priority order when submitting the training or processing job. For example, your list might contain a preference of two instances of ml.g6.48xlarge or four instances of ml.g5.48xlarge. SageMaker works through the list and launches your job on the first configuration where capacity is available. You can also configure the capacity sourcing from on-demand sources or from your reserved SageMaker Flexible Training Plans within the same job submission. This feature simplifies the process of getting compute for your jobs during high-demand periods and reduces the undifferentiated manual retrying you would otherwise do, all within the SageMaker training and processing job APIs you already use. Instance preference lists for SageMaker training and processing jobs is available today in all AWS Regions where SageMaker is available through the SageMaker CLIs, APIs, SDKs and Console UI. To learn more, see our https://docs.aws.amazon.com/sagemaker/latest/dg/train-instance-preferences.html or our https://aws.amazon.com/blogs/machine-learning/announcing-instance-preference-lists-for-amazon-sagemaker-ai-training-jobs/.

Today, Amazon SageMaker AI announces instance preference lists for training and processing jobs, making it easier and faster to find compute capacity for your workloads. Many AI training, fine-tuning, and data processing workloads run comparably well on any of several instance types or sizes. However, before now, you had to name only one instance type at the time of job submission and wait for SageMaker to find that specific instance for your job. For high-demand GPUs during peak periods, where wait times can be unpredictable, customers sometimes had to build complex retry logic or concurrently submit multiple jobs with different instance types to find the first available option. Now you can simply provide a prioritized list of the instance types your workload accepts, and SageMaker automatically runs your job on the first available configuration from your preferences. With this solution, your training or processing job will likely start sooner. To use this feature, you specify your instance type and count preferences in priority order when submitting the training or processing job. For example, your list might contain a preference of two instances of ml.g6.48xlarge or four instances of ml.g5.48xlarge. SageMaker works through the list and launches your job on the first configuration where capacity is available. You can also configure the capacity sourcing from on-demand sources or from your reserved SageMaker Flexible Training Plans within the same job submission. This feature simplifies the process of getting compute for your jobs during high-demand periods and reduces the undifferentiated manual retrying you would otherwise do, all within the SageMaker training and processing job APIs you already use. Instance preference lists for SageMaker training and processing jobs is available today in all AWS Regions where SageMaker is available through the SageMaker CLIs, APIs, SDKs and Console UI. To learn more, see our documentation or our launch blog.

Amazon SageMaker Studio quick setup now completes in under twenty seconds, reduced from over two minutes. Whether you are building ML pipelines, exploring data, developing with notebooks, or fine-tuning foundation models, you can go from sign-in to a fully configured Studio environment almost instantly. As part of this streamlined setup, newly created Studio environments now come with serverless model customization permissions automatically configured. A new managed policy, AmazonSageMakerModelCustomizationCoreAccess, is created and attached for you, providing permissions for serverless model customization jobs including fine-tuning with custom reward functions for reinforcement learning, model evaluation, and deployment to SageMaker or Bedrock endpoints. This eliminates the need to manually create and configure IAM roles and policies before you can start experimenting. For existing Studio environments, actionable messages with direct links to documentation guide you through adding these permissions. This feature is available in all AWS Commercial Regions where Amazon SageMaker Studio is supported. To get started, create a new Studio environment using quick setup in the SageMaker AI Console. To learn more, see https://docs.aws.amazon.com/sagemaker/latest/dg/onboard-quick-start.html and https://docs.aws.amazon.com/sagemaker/latest/dg/model-customize-open-weight-prereq.html in the Amazon SageMaker documentation.

Amazon SageMaker Studio now supports direct integration from Hugging Face, letting you go from discovering a model to working with it inside a fully configured Studio environment in a single click. Select any supported model on Hugging Face and choose "Customize on SageMaker AI" or "Deploy on SageMaker AI" to land directly on the corresponding workflow page with the model pre-loaded and ready to use. Previously, getting from model discovery to a working environment required navigating the AWS Console to find SageMaker AI, configuring an environment, setting up IAM permissions for serverless model customization, and in many cases requesting GPU quota increases through Service Quotas before running a first job. Now, new customers complete a standard AWS sign-up and receive a SageMaker Studio environment created in seconds with pre-configured permissions for serverless model customization jobs including fine-tuning with custom reward functions for reinforcement learning, model evaluation, and deployment to SageMaker or Bedrock endpoints. Verified customers receive default GPU access to G5, G6, and G4dn instances across endpoint deployments, training jobs, and notebooks without requesting quota increases, and quota limit and utilization information is visible for each instance type directly inside the Studio environment. Returning customers signing in from Hugging Face or SageMaker product pages select their environment and land directly inside SageMaker Studio with the model ready to use. This feature is available in all AWS Commercial Regions where Amazon SageMaker Studio is supported. To get started, visit any supported model on https://huggingface.co/models and select "Customize on SageMaker AI" or "Deploy on SageMaker AI," or click Get Started from the https://aws.amazon.com/sagemaker/ai/studio/ page. To learn more, see https://docs.aws.amazon.com/sagemaker/latest/dg/studio-updated-launch.html in the Amazon SageMaker documentation.

Today, https://docs.aws.amazon.com/sagemaker/latest/dg/how-it-works-training.html announces the new Amazon EC2 P5 instance size with one NVIDIA H100 GPU. It allows businesses to right-size their machine learning (ML) and high-performance computing (HPC) resources with cost-effectiveness. The new P5 instance type gives customers the flexibility to begin with smaller configurations and expand incrementally with fine-grained control, delivering enhanced cost management for their infrastructure investments. P5.4xlarge instances are now available through https://docs.aws.amazon.com/sagemaker/latest/dg/reserve-capacity-with-training-plans.html in the following AWS Regions: US East (North Virginia, Ohio), US West (Oregon), Europe (London), Asia Pacific (Mumbai, Sydney, Tokyo), and South America (Sao Paulo). Additionally, the instance type can be purchased through SageMaker On-Demand and Spot in the Europe (London), Asia Pacific (Mumbai, Jakarta, Tokyo), and South America (Sao Paulo) AWS Regions. To learn more about P5.4xlarge instances, visit https://aws.amazon.com/ec2/instance-types/p5/.

Amazon SageMaker AI now offers multi-turn reinforcement learning (RL), a new serverless model customization technique for fine-tuning models on multi-step, agentic tasks. SageMaker AI model customization lets you adapt foundation models using techniques such as supervised fine-tuning, reinforcement learning from verifiable rewards (RLVR), and reinforcement learning from AI feedback (RLAIF), without the undifferentiated heavy lifting of building and operating your own training infrastructure. Multi-turn RL extends this by training models against your own agent environment and rewarding the full sequence of decisions an agent makes across a task, helping you specialize smaller, lower-cost models to match or exceed the task accuracy of larger general-purpose models on your target workload. Training models that power agents to reliably complete multi-step tasks is complex and time-consuming, often requiring custom infrastructure that takes weeks to build. SageMaker's Multi-turn RL offering handles this for you. You can connect your agent running on Amazon Bedrock AgentCore Runtime for fully managed hosting, or on Amazon EKS, Amazon EC2, AWS Fargate, or any infrastructure using the framework of your choice. SageMaker AI manages the full training loop, from rollout orchestration and trajectory collection to training and checkpoint management. Built-in MLflow tracking lets you inspect agent trajectories, rewards, and traces. Evaluation jobs report reward, pass@k, and trajectory metrics so you can benchmark a model before deploying it to a SageMaker AI endpoint or Amazon Bedrock. Multi-turn RL runs as a fully serverless capability, so you pay only for the tokens processed, with no infrastructure to provision or manage. Multi-turn RL is available today through SageMaker Studio and the SageMaker Python SDK as part of Amazon SageMaker AI model customization. Supported models include Qwen 3.6 27B, Nova Lite 2.0, GPT-OSS-20B and Gemma 31B in us-west-2, and Nova Lite 2.0, GPT-OSS-20B in us-east-1. To get started with multi-turn reinforcement learning in SageMaker AI, visit the https://docs.aws.amazon.com/sagemaker/latest/dg/model-customize-mtrl.html. 

Amazon SageMaker AI now offers multi-turn reinforcement learning (RL), a new serverless model customization technique for fine-tuning models on multi-step, agentic tasks. SageMaker AI model customization lets you adapt foundation models using techniques such as supervised fine-tuning, reinforcement learning from verifiable rewards (RLVR), and reinforcement learning from AI feedback (RLAIF), without the undifferentiated heavy lifting of building and operating your own training infrastructure. Multi-turn RL extends this by training models against your own agent environment and rewarding the full sequence of decisions an agent makes across a task, helping you specialize smaller, lower-cost models to match or exceed the task accuracy of larger general-purpose models on your target workload. Training models that power agents to reliably complete multi-step tasks is complex and time-consuming, often requiring custom infrastructure that takes weeks to build. SageMaker's Multi-turn RL offering handles this for you. You can connect your agent running on Amazon Bedrock AgentCore Runtime for fully managed hosting, or on Amazon EKS, Amazon EC2, AWS Fargate, or any infrastructure using the framework of your choice. SageMaker AI manages the full training loop, from rollout orchestration and trajectory collection to training and checkpoint management. Built-in MLflow tracking lets you inspect agent trajectories, rewards, and traces. Evaluation jobs report reward, pass@k, and trajectory metrics so you can benchmark a model before deploying it to a SageMaker AI endpoint or Amazon Bedrock. Multi-turn RL runs as a fully serverless capability, so you pay only for the tokens processed, with no infrastructure to provision or manage. Multi-turn RL is available today through SageMaker Studio and the SageMaker Python SDK as part of Amazon SageMaker AI model customization. Supported models include Qwen 3.6 27B, Nova Lite 2.0, GPT-OSS-20B and Gemma 31B in us-west-2, and Nova Lite 2.0, GPT-OSS-20B in us-east-1. To get started with multi-turn reinforcement learning in SageMaker AI, visit the Amazon SageMaker AI documentation.

Amazon SageMaker Studio quick setup now completes in under twenty seconds, reduced from over two minutes. Whether you are building ML pipelines, exploring data, developing with notebooks, or fine-tuning foundation models, you can go from sign-in to a fully configured Studio environment almost instantly. As part of this streamlined setup, newly created Studio environments now come with serverless model customization permissions automatically configured. A new managed policy, AmazonSageMakerModelCustomizationCoreAccess, is created and attached for you, providing permissions for serverless model customization jobs including fine-tuning with custom reward functions for reinforcement learning, model evaluation, and deployment to SageMaker or Bedrock endpoints. This eliminates the need to manually create and configure IAM roles and policies before you can start experimenting. For existing Studio environments, actionable messages with direct links to documentation guide you through adding these permissions. This feature is available in all AWS Commercial Regions where Amazon SageMaker Studio is supported. To get started, create a new Studio environment using quick setup in the SageMaker AI Console. To learn more, see Quick setup and Model Customization permissions setup in the Amazon SageMaker documentation.

Today, SageMaker Training and Processing Jobs announces the new Amazon EC2 P5 instance size with one NVIDIA H100 GPU. It allows businesses to right-size their machine learning (ML) and high-performance computing (HPC) resources with cost-effectiveness. The new P5 instance type gives customers the flexibility to begin with smaller configurations and expand incrementally with fine-grained control, delivering enhanced cost management for their infrastructure investments. P5.4xlarge instances are now available through SageMaker Flexible Training Plans in the following AWS Regions: US East (North Virginia, Ohio), US West (Oregon), Europe (London), Asia Pacific (Mumbai, Sydney, Tokyo), and South America (Sao Paulo). Additionally, the instance type can be purchased through SageMaker On-Demand and Spot in the Europe (London), Asia Pacific (Mumbai, Jakarta, Tokyo), and South America (Sao Paulo) AWS Regions. To learn more about P5.4xlarge instances, visit Amazon EC2 P5 instances.