Amazon Bedrock adds ability to import custom models

Amazon has announced several major updates to Amazon Bedrock, its platform that provides developers with access to foundation models (FMs) from AI companies such as AI21 Labs, Anthropic, Cohere, Meta, Mistral AI, Stability AI, and Amazon. 

Amazon Bedrock is a managed service, handling the underlying infrastructure for these models so that developers can build generative AI applications without worrying about managing that. 

The company added a new Custom Model Import option, which allows developers to add their own proprietary models to Amazon Bedrock. This allows them to leverage domain-specific models that were trained on their own data. 

Companies can create their models in Amazon SageMaker or another tool and then import them into Amazon Bedrock.

Imported models go through an automated validation process to ensure that they meet proper safeguards and responsible AI standards. It currently supports three open model architectures — Flan-T5, Llama, and Mistral — and more will be added in the future. 

Amazon also added a new Model Evaluation capability, which allows teams to compare different models within Amazon Bedrock. According to Amazon, companies can spend days or weeks evaluating different models to determine if they meet their use case. The goal of this new feature is to significantly cut down on evaluation time. 

Developers can select evaluation criteria, such as accuracy or robustness, and upload their own dataset. Then, Amazon Bedrock runs evaluations on different models and generates a report that shows how each model performed for each criteria. 

Another new feature is Guardrails for Amazon Bedrock, which makes it easier for companies to implement safeguards that are tailored to their own responsible AI policies. Most models have their own safeguards, but this acts as an additional layer of safety on top and can block 85% of harmful content, Amazon says.

Developers can create guardrails using a natural language description and can add additional filters for hate speech, insults, sexualized language, prompt injection, and violence. They can also add filters to remove personal information, specific words, or profanity. 

There were also a couple of updates related to Amazon Titan, which is a family of FMs created by Amazon. Amazon Titan Text Embeddings V2 is an FM that is optimized for retrieval-augmented generation (RAG), and it will be available next week,  exclusively on Amazon Bedrock. 

Amazon Titan Image Generator allows developers to enhance or edit existing images via natural language prompts, and is generally available now. 

“With today’s announcements, we continue to innovate rapidly for our customers by doubling-down on our commitment to provide them with the most comprehensive set of capabilities and choice of industry-leading models, further democratizing generative AI innovation at scale,” said Dr. Swami Sivasubramanian, vice president of AI and Data at AWS.