Amazon Web Services is releasing new capabilities for its Bedrock artificial intelligence service, including the ability for customers to import their own custom AI models to run alongside models from major providers.
It’s part of a larger effort to appeal to AI application developers, expanding the capabilities of its cloud platform by offering access to a variety of AI models through the Bedrock managed service. Amazon is competing against other cloud giants such as Microsoft and Google, and newer AI platforms such as OpenAI and HuggingFace.
Custom Model Import “allows customers that have spent a bunch of time creating something outside of Bedrock to bring them inside and participate as a first-class model with the rest of the Bedrock workflows,” said Vasi Philomin, the AWS vice president of generative AI, in an interview this week.
The new capability will be available starting today in preview, the company said. Early users include Amazon’s internal Rufus team, which is making the model that powers its generative AI shopping chatbot available to other Amazon developers internally via Bedrock using the custom model feature.
More from AWS Bedrock this morning:
Updates to Amazon’s Titan foundation models: The company is launching its Titan Text Embeddings Model V2 next week, improving accuracy and efficiency; and making its Titan Image Generator generally available today, including built-in invisible watermarking for identifying AI-generated images.
More third-party models: Meta’s Llama 3 foundation models, for building and experimenting with generative AI apps, are now available on Bedrock. Cohere’s Command R and Command R+ models for enterprise AI applications will be available soon. This follows the recent release of Anthopic’s Claude 3 Opus on Bedrock.
Now generally available: The company is broadly releasing its previously announced Model Evaluation capability for Bedrock to help developers compare different models; and its Guardrails feature, which incorporates safeguards to remove or block personal and sensitive information, profanity, and harmful content.
