OpenAI Open-Weight Models Now Live on AWS Bedrock & SageMaker

Amazon Web Services (AWS) has expanded its selection of foundation models by making two new OpenAI open-weight models, gpt-oss-120b and gpt-oss-20b, available on Amazon Bedrock and Amazon SageMaker JumpStart. This integration provides developers and organizations with enhanced options for building AI applications, offering greater control over their infrastructure and data.

These OpenAI models are designed for a range of text generation and reasoning tasks. They demonstrate strong performance in areas such as coding, scientific analysis, and mathematical reasoning, comparable to other leading models in the industry. Both models feature a substantial 128K context window and offer adjustable reasoning levels (low, medium, or high) to align with specific application requirements. Their capabilities can be extended through the use of external tools and they support integration into agentic workflows, for example, using frameworks like Strands Agents.

The availability of these open-weight models on AWS platforms underscores a commitment to offering a diverse array of advanced foundation models from various AI innovators. This comprehensive selection aims to provide users with the flexibility to choose the most suitable model for their AI workloads.

Through Amazon Bedrock, users can seamlessly experiment with different models, combine capabilities, and switch between providers without the need for extensive code rewrites. This flexibility transforms model choice into a strategic advantage, allowing organizations to continuously adapt their AI strategy as new innovations emerge. At launch, these new OpenAI models are accessible in Bedrock via an OpenAI-compatible endpoint, enabling integration with the OpenAI SDK or direct use with Bedrock’s InvokeModel and Converse APIs.

For those utilizing Amazon SageMaker JumpStart, the platform facilitates the rapid evaluation, comparison, and customization of models for specific use cases. Users can then deploy either the original or a fine-tuned version of the model into production environments using the SageMaker AI console or the SageMaker Python SDK.

Key Information for Users:

• Regional Availability: The new OpenAI open-weight models are currently available in Amazon Bedrock in the US West (Oregon) AWS Region. For Amazon SageMaker JumpStart, these models are supported in US East (Ohio, N. Virginia) and Asia Pacific (Mumbai, Tokyo) regions.

• Transparency and Customization: Each model is equipped with full chain-of-thought output capabilities, providing detailed insight into the model’s reasoning process. This transparency is particularly beneficial for applications requiring high levels of interpretability. The open-weight nature of these models allows users to modify, adapt, and customize them to their specific needs, enabling fine-tuning for unique use cases, integration into existing workflows, and the development of new, specialized models.

• Security and Compatibility: Security and safety measures are integrated into the core design of these models, supported by comprehensive evaluation processes. The models maintain compatibility with the standard GPT-4 tokenizer.

• Deployment Flexibility: Users have the option to deploy these models in their preferred environment: either through the serverless experience offered by Amazon Bedrock or leveraging the extensive machine learning development capabilities of SageMaker JumpStart. Information regarding the costs associated with using these models and services is available on the respective Amazon Bedrock and Amazon SageMaker AI pricing pages.

Developers and organizations can begin utilizing these OpenAI open-weight models on AWS via the Amazon Bedrock console or the Amazon SageMaker AI console.