AWS continues to solidify its position as a leader in the AI landscape by introducing two new OpenAI models with open weights, now available through Amazon Bedrock and Amazon SageMaker JumpStart. The gpt-oss-120b and gpt-oss-20b models are tailored for text generation and reasoning tasks, providing developers and organizations with enhanced options for building AI applications while maintaining complete control over their infrastructure and data.
Model Capabilities and Features
These newly launched open weight models are particularly adept at coding, scientific analysis, and mathematical reasoning, boasting performance levels that rival leading alternatives in the market. Both models support a context window of 128K and offer adjustable reasoning levels—low, medium, and high—to align with specific use case requirements. Additionally, they can integrate with external tools to further enhance their functionality, making them suitable for use in agentic workflows, such as those utilizing frameworks like Strands Agents.
With Amazon Bedrock and Amazon SageMaker JumpStart, AWS empowers users with the flexibility to innovate by accessing a diverse array of foundation models from top AI companies, including the newly introduced OpenAI models. This extensive selection allows businesses to align their AI workloads with the most appropriate models, ensuring optimal performance and results.
Amazon Bedrock facilitates seamless experimentation with various models, enabling users to mix and match capabilities and switch between providers without the need to rewrite code. This adaptability transforms model selection into a strategic advantage, allowing organizations to evolve their AI strategies in tandem with emerging innovations. The new models can be accessed via an OpenAI-compatible endpoint in Bedrock, where users can utilize the OpenAI SDK or the Bedrock InvokeModel and Converse API for integration.
Practical Implementation
To get started with the OpenAI open weight models in Amazon Bedrock, users can navigate to the Amazon Bedrock console, select Model access from the Configure and learn section, and request access to the listed OpenAI models. Once access is granted, the Chat/Test playground can be used to evaluate the models. For instance, selecting the gpt-oss-120b model allows users to input prompts and receive outputs that detail the reasoning process behind the generated results.
For those utilizing the OpenAI SDK, configuration of the API endpoint and authentication via an Amazon Bedrock API key is straightforward. Users can set environment variables to specify the AWS Region endpoint and invoke the model using the OpenAI Python SDK. The process involves writing a simple script to interact with the model, enabling users to explore its capabilities effectively.
Deployment and Customization
Building an AI agent is also a possibility, with users able to select any framework that supports the Amazon Bedrock API or the OpenAI API. For example, Strands Agents can be employed to create a basic agent using the Amazon Bedrock API, allowing for local testing before deploying in a production environment. Amazon Bedrock AgentCore offers a fully managed serverless runtime, along with memory and identity management, simplifying the deployment process.
In the Amazon SageMaker AI console, users can access the OpenAI open weight models through SageMaker Studio. Initial setup requires establishing a SageMaker domain, which can be configured for either a single user or an organization. After deploying a selected model, users can invoke it in SageMaker Studio or via any AWS SDKs, streamlining the integration process.
Key Considerations
The new OpenAI open weight models are currently available in the US West (Oregon) AWS Region through Amazon Bedrock, while SageMaker JumpStart supports them in the US East (Ohio, N. Virginia) and Asia Pacific (Mumbai, Tokyo) regions. Each model features full chain-of-thought output capabilities, offering valuable insights into the model’s reasoning process, which is essential for applications requiring high interpretability and validation.
These models provide the flexibility to modify, adapt, and customize them according to specific needs, allowing businesses to fine-tune their applications and integrate them into existing workflows. Security and safety measures are integral to these models, ensuring robust evaluation processes are in place. Compatibility with the standard GPT-4 tokenizer further enhances their usability.
Whether through the serverless experience of Amazon Bedrock or the comprehensive machine learning capabilities of SageMaker JumpStart, users can leverage these models in their preferred environments. For detailed information on associated costs, users are encouraged to visit the Amazon Bedrock pricing and Amazon SageMaker AI pricing pages.
To explore the parameters for the models and the chat completions API, refer to the Amazon Bedrock documentation. The opportunity to harness OpenAI open weight models on AWS is now at your fingertips, inviting you to start your journey in the Amazon Bedrock console or the Amazon SageMaker AI console.
