{
  "id": 16082945,
  "title": "Announcing new ways to customize models with SageMaker AI, without managing GPUs",
  "url": "https://urgent.news/2026/10/22/announcing-new-ways-to-customize-models-with-sagemaker-ai-without",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-22T18:00:04.000Z",
  "source": {
    "name": "AWS Machine Learning",
    "slug": "aws-machine-learning",
    "url": "https://aws.amazon.com/blogs/machine-learning/announcing-new-ways-to-customize-models-with-sagemaker-ai-without-managing-gpus/"
  },
  "original_language": "en",
  "account": "Amazon SageMaker AI has expanded its capabilities to help companies customize their AI models without the need to manage GPUs. Traditionally, building customized models required selecting the right technique, setting up GPU infrastructure, and creating multi-turn reinforcement learning environments. However, with the general availability of training sessions, SageMaker AI now provides a serverless model customization experience.\n\nWith training sessions, users can control post-training logic without managing infrastructure. They write and coordinate the training loop, including custom loss functions or reward signals. SageMaker AI takes care of GPU allocation, distributed execution, fault recovery, and checkpoint storage.\n\nCustomers can choose various post-training techniques such as SFT, DPO, RLVR, RLAIF, or multi-turn RL. They can also decide how much of the model to update using Low-Rank Adaptation (LoRA) or full fine-tuning, and chain techniques together through continuous customization.\n\nTo simplify the process, Amazon SageMaker AI offers a cookbook with examples and use cases. Users describe their use case, and the agent handles data preparation, technique selection, training, evaluation, and deployment. The agent works within popular environments like SageMaker Studio, Kiro, Claude Code, Codex, or GitHub Copilot.\n\nThe training sessions are particularly useful for multi-turn reinforcement learning, where agents need multiple interactions to achieve the desired outcome. For instance, a travel booking agent would require several rounds of interaction before determining if the correct outcome is reached. SageMaker AI handles the GPU-intensive work by running the training loop and reward function in the user's environment. The sampler generates responses, the trainer calculates gradients, and updates the model weights.\n\nEach session utilizes an AWS IAM execution role, allowing users to add their own AWS KMS key and VPC configuration. Data remains in the user's Amazon S3 buckets, and checkpoints are stored there as well. SageMaker AI does not use the data or customized weights to train other models.\n\nAfter deployment, users can evaluate the model for accuracy, bias, and safety, and apply guardrails and monitor performance using Amazon CloudWatch. This streamlined model customization process enables companies to build accurate, cost-effective, and tailored AI solutions without the hassle of managing GPU infrastructure.",
  "summary": "Amazon SageMaker AI serverless model customization now includes training sessions, generally available today. You write and coordinate your own post-training loop (custom loss functions, reward signals, or multi-turn reinforcement learning) while SageMaker AI handles GPU allocation, distributed execution, fault recovery, and checkpoint storage.",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}