Amazon Titan

What Titan is

Amazon Titan is a family of Amazon's own foundation models for Amazon Bedrock. It was first announced on 13 April 2023 with two kinds of model: a generative large language model and an embeddings model.^[1] The family later grew to include text models and image and multimodal models. In use, Titan models could be used as they were or privately customised in Bedrock.^[3] Titan models can also be fine-tuned on customer data in Bedrock.^[3] Amazon did not publish parameter counts or architecture details for the Titan models.^[1][3]

Release and model types

The first Titan model to become generally available was Amazon Titan Embeddings, on 28 September 2023. It supported more than 25 languages and inputs of up to 8,000 tokens.^[2] On 29 November 2023, two text models became generally available: Titan Text Express, with a context length of up to 8,192 tokens, and Titan Text Lite, a lower-cost English model with a context length of 4,096 tokens.^[3] The text models launched in US East (N. Virginia), US West (Oregon), Singapore, Tokyo and Frankfurt.^[3]

The two text models served different needs within the family. Express offered a longer context window, while Lite was positioned as a lower-cost English model with a shorter context window.^[3] Both were part of the same Titan family and both were available in the listed regions.^[3] The option to fine-tune on customer data in Bedrock meant that users could adapt a Titan model to their own material; AWS also described the models as usable as-is or privately customised.^[3]

Because Amazon did not publish parameter counts or architecture details, the public description of Titan focuses on what the models can do and how they are deployed: their context lengths, supported languages, embedding dimensions and the regions where they are available.^[1][3]

Images and multimodal embeddings

At the same time as the text model general availability, Amazon previewed Titan Image Generator. It creates and edits images from English prompts, including inpainting and outpainting, and adds an invisible watermark to every image.^[3] AWS said the image model used a built-in segmentation model to edit images from text prompts, supporting inpainting with an image mask and outpainting to extend or change a background.^[3] It could be customised with proprietary images, such as a previous marketing campaign, to follow brand guidelines or match a style.^[3]

Amazon also launched Titan Multimodal Embeddings on the same day. It converts images and short English text into embeddings of 1,024 dimensions by default.^[3] An embedding is a numeric representation of content that software can compare or search over. For this model, the inputs are images and short English text, and the output is a 1,024-dimension embedding by default; applications can use such embeddings to compare or organise content.^[3]

Position in Amazon's model line

Titan was part of Amazon's model offering through Bedrock. AWS said Titan models could be used as-is or privately customised, and that Titan Text Embeddings was available alongside Lite, Express and Image Generator.^[3] The family was succeeded as Amazon's main model family by Amazon Nova, introduced on 3 December 2024.^[4] The Titan name therefore marks an early Amazon model family for Bedrock, covering text, embeddings and image generation before Nova took over as the main family.