AWS unveils new generative AI models

Amazon Web Services (AWS) has unveiled its Nova family of generative artificial intelligence (GenAI) foundation models that promise faster speeds, lower costs and broader capabilities than competing offerings. 

Announced by Amazon CEO Andy Jassy at AWS re:Invent 2024 in Las Vegas, the Nova family comprises frontier models that cater to different needs. These include Nova Micro, a text-only model optimised for speed and cost, along with three multimodal models: Nova Lite, Nova Pro and Nova Premier.

The multimodal models can process text, images and video, generating sophisticated text outputs. While Nova Micro, Lite and Pro are available immediately, the more powerful Nova Premier model that can handle complex reasoning tasks is slated for release in the first quarter of 2025.

Citing benchmarks comparing Nova models to leading competitors like Google’s Gemini and Meta’s Llama, Jassy claimed that Nova Micro performs equally or better than Gemini 1.5 Flash-8B and Llama 3.1 8B on standard benchmarks.

Similarly, Nova Lite showed comparable or superior performance compared to OpenAI’s GPT-4o Mini and Gemini 1.5 Flash. The forthcoming Nova Premier, positioned as Amazon’s most powerful multimodal model, is expected to further elevate the bar.

Cost-effectiveness was another key selling point. Jassy said the Nova models are about 75% less expensive than comparable models, making them a more accessible option for developers. He also touted their speed, claiming Nova models offer the lowest latency currently available. This enhanced performance is further bolstered by Amazon's purpose-built inference chip, Inferentia2, which powers its latency-optimised inf2 compute instances.

The Nova models are integrated with Amazon Bedrock, a fully managed service for building and scaling generative AI applications. The integration is expected to unlock capabilities such as fine-tuning models with specific data, grounding responses with factual information, and orchestrating AI agents to automate workflows.

Besides Nova Micro, Lite, Pro and Premier, AWS has also introduced two Nova models for content creators: Nova Canvas for image generation and Nova Reel for video generation.

Nova Canvas allows users to create studio-quality images from natural language prompts, offering editing capabilities and built-in safeguards for responsible AI usage, including watermarking and content moderation. When benchmarked against the DALL-E 3 model from OpenAI and Stability AI’s Stable Diffusion 3.5 model, Canvas reportedly outperforms both in image quality and instruction-following.

Nova Reel, Amazon’s video generation model, lets users create videos with granular control over camera movements, zoom and panning. Initially supporting six-second videos, Reel will soon expand to two-minute videos, catering to marketing and advertising needs. Human evaluations suggest Reel surpasses other video generation services in quality and performance.

Looking ahead, Jassy teased upcoming additions to the Nova family, including a speech-to-speech model expected in the first quarter next year and a groundbreaking “any-to-any” model by mid-year. The latter model will be capable of processing and generating text, speech, images and video, representing a significant leap forward in multimodal AI capabilities.

Even as AWS is investing in its own frontier models like Nova, Jassy said the company will continue to partner with other model providers, enabling customers to use models that best suit their needs based on factors like performance, cost and specialised capabilities.

“The reality is that all of you are going to use different models for different reasons at different times,” he said. “At AWS, we are going to give you the very best combination of all these as we always do.”

The launch of the Nova models signals Amazon’s aggressive push to improve its foundation models, challenging rivals and offering developers a new toolkit for innovation. With its focus on practicality at a lower cost and integration with AWS services, Nova has the potential to reshape AI application development.