Amazon Launches Its Own Generative AI Foundation Models

LAS VEGAS — Amazon Web Services entered the Generative AI foundation model arena Tuesday, introducing four foundation models at its Re:Invent developer conference. Previously, Amazon relied on providing access to other models, such as Anthropic’s Claude.

Amazon CEO and President Andy Jassy, who previously lead AWS, made a guest appearance during AWS President Matt Garman’s keynote to announce the new models.

“They’re about 75% less expensive than the other data models,” Jassy told the audience, adding that they are fast models with respect to latency as well.

They’re accessible through Bedrock, which is AWS’s managed service that provides access to a variety of foundation models from AI companies through a single API.

These new models are optimized to be effective in AI agentic applications that might require interacting with an organization’s property system and data via APIs, the company said.

The models aren’t just available in Bedrock — they are integrated into Bedrock’s features to support fine-tuning the models, Jassy said.

“It’s deeply integrated in Bedrock’s knowledge bases, so that you can use RAG to ground your answers in your own data,” he said. “Then also, we’ve optimized these models to work with the proprietary systems’ APIs, so that you can actually do multiple orchestrated, automatic steps [for] agentic behavior much more easily with these models.”

The Models at a Glance

The models were introduced by Jassy in order of the smallest to largest and cost-effectiveness:

• Nova Micro, a fast text-to-text model that offers the lowest latency responses of the four at a low cost. “It’s laser fast, very cost-effective,” Jassy said of the model.
• Nova Lite, a low-cost multimodal modal that is lightning-fast for processing image, video and text inputs.
• Nova Pro, a multimodal model with a combination for accuracy, speed and cost designed for a wide range of tasks;
• Nova Premier, a multimodal model that processes text, images and videos. It will be available the first quarter of 2025 and is the most capable for complex reasoning tasks. It can also be used as a teacher to distill custom models that can process text, images, and videos to generate text.

Amazon benchmarked the Nova LLMs against competitor LLMs and the models were “very competitive,” Jassy said.

“If you compare the light to OpenAI’s GPT-4o mini, you can see that we’re equal or better on 17 of 19 benchmarks,” he said. ”Then if you look at the pro, same story if you compare it to GBT-4o — it’s equal or better on 17 of the 20 benchmarks, [and] it’s equal or better on 16 of the 21 benchmarks versus Gemini.”

It also launched two additional AI models: Amazon Nova Canvas, which generates studio-quality images from prompts; and Amazon Nova Reel, a video generation model that allows customers to create video from text and images. Currently, Nova Reel can generate six-second videos.

Amazon also plans to introduce a speech-to-speech model in the first quarter of 2024. It’s designed to transform conversational AI applications by understanding speech input in natural language, interpreting verbal and non-verbal cues, and delivering natural back-and-forth interactions with low latency.

Editor’s Note: Updated Dec. 6 to correct Mr. Garman’s name.