Amazon SageMaker AI is a cloud-based machine-learning platform that allows developers to create, train, and deploy machine-learning (ML) models on the cloud. Launched in November 2017 at the AWS re:Invent conference, it supports deployment of ML models on embedded systems and edge devices. The platform is part of Amazon Web Services (AWS) and integrates with other AWS services for data storage, batch processing, and real-time data streams.
SageMaker provides multiple levels of abstraction for developers, ranging from pre-trained models to fully custom algorithms. It includes built-in ML algorithms, managed instances of TensorFlow and Apache MXNet, and interfaces for various programming languages. The platform has evolved with features such as automatic scaling, reinforcement learning, and edge deployment capabilities.
Capabilities
SageMaker enables developers to operate at different levels of abstraction when training and deploying machine-learning models. At its highest level, it provides pre-trained ML models that can be deployed as-is. It also offers a number of built-in ML algorithms that developers can train on their own data, covering common tasks like classification, regression, and clustering.
The platform features managed instances of TensorFlow and Apache MXNet, where developers can create their own ML algorithms from scratch. This flexibility supports both machine learning and deep learning workflows. Developers can connect SageMaker-enabled ML models to other AWS services, such as Amazon DynamoDB for structured data storage, AWS Batch for offline batch processing, or Amazon Kinesis for real-time processing.
Development Interfaces
Various interfaces are available for developers to interact with SageMaker. A web API can remotely control a SageMaker server instance, and while the API is agnostic to the programming language used by the developer, Amazon provides SageMaker API bindings for several languages, including Python, JavaScript, Ruby, Java, and Go. The SageMaker Python SDK is accessible through Anaconda's community-led conda-forge channel.
SageMaker also provides managed Jupyter Notebook instances for interactively programming SageMaker and other applications. In December 2025, new serverless customization in SageMaker AI was announced to interface with AI models like Amazon Nova, Llama, Qwen, DeepSeek, and GPT-OSS, expanding its support for generative AI workloads.
History and Features
SageMaker was launched on 2017-11-29 at the AWS re:Invent conference. Early development focused on expanding training and inference capabilities:
- 2018-02-27: Managed TensorFlow and MXNet deep neural network training and inference became supported within SageMaker.
- 2018-02-28: SageMaker automatically scales model inference to multiple server instances.
- 2018-07-13: Support was added for recurrent neural network training, word2vec training, multi-class linear learner training, and distributed deep neural network training in Chainer with Layer-wise Adaptive Rate Scaling (LARS).
- 2018-07-17: AWS Batch Transform enabled high-throughput non-real-time machine-learning inference in SageMaker.
- 2018-11-08: Support for training and inference of Object2Vec word embeddings was introduced.
- 2018-11-27: SageMaker Ground Truth was launched, making it easier for developers to label their data using human annotators through Mechanical Turk, third-party vendors, or their own employees.
- 2018-11-28: SageMaker Reinforcement Learning (RL) enabled developers and data scientists to develop reinforcement learning models at scale.
- 2018-11-28: SageMaker Neo enabled deep neural network models to be deployed from SageMaker to edge devices such as smartphones and smart cameras.
- 2018-11-29: The AWS Marketplace for SageMaker was launched, allowing third-party developers to buy and sell machine-learning models that can be trained and deployed in SageMaker.
- 2019-01-27: SageMaker Neo was released as open-source software.
Integration and Ecosystem
SageMaker integrates with the broader AWS ecosystem, allowing developers to build end-to-end ML pipelines. It supports data augmentation and other preprocessing techniques, and its managed infrastructure reduces the operational overhead of training and deploying models. The platform's support for edge deployment through SageMaker Neo addresses scenarios where low latency or offline operation is required, such as in autonomous vehicles or automotive systems.
SageMaker's evolution reflects broader trends in artificial intelligence, including the shift toward large language models and serverless computing. The December 2025 announcement of serverless customization for models like Llama and DeepSeek indicates a move toward more flexible and cost-effective deployment options for generative AI applications.