What Amazon Nova Premier reveals about AI models and agents
While the model is not unique, the vendor said it is the fastest and most cost-effective model in Amazon Bedrock at its intelligence level, which might appeal to some enterprises.
A few months after introducing its new family of generative AI models, Amazon Nova, AWS revealed that Amazon Nova Premier is now generally available.
The cloud giant said Nova Premier is now available on the Amazon Bedrock model platform, like the rest of its Nova family of models, including Nova Pro, Lite and Micro. The foundation model is best for complex tasks and is a teacher for model distillation. In model distillation, the teacher model is used to fine-tune a smaller model.
With Nova Premier and the Amazon Bedrock Model Distillation tool, users can create cost-effective and low-latency versions of Nova Pro, Lite and Micro, AWS said.
Like the other Nova family models, Nova Premier can process input text, images and videos. The vendor said the model is good at tasks that require a deep understanding of context, multistep planning and precise execution.
AWS said that while Nova Premier is not a reasoning model, it can be used for multi-agent collaboration applications.
Not unique
Amazon Nova Premier is an example of how AI models continue growing and evolving. However, analysts said the Nova family of models is not necessarily unique compared with other foundation models.
They are a significant improvement from the Titan family of models, but I've not yet seen them emerge as a particular standout in any vertical or horizontal use case.
Rowan CurranAnalyst, Forrester Research
"They are a significant improvement from the Titan family of models, but I've not yet seen them emerge as a particular standout in any vertical or horizontal use case," said Rowan Curran, an analyst at Forrester Research. Titan is an older Amazon model family.
He added that while AWS highlights that Nova Premier is good at distillation, there's nothing particularly special about it compared with other models on the market. However, enterprises might be interested in Nova Premier because they have access to an intelligent model from a trusted provider like AWS that they can use to create small, distilled models for their applications.
Enterprises might also be interested in Nova Premier because of the cost and speed claims AWS makes. The cloud provider said Nova Premier is the fastest and most cost-effective model available in Amazon Bedrock.
"That's the thing that could be combined with distillation, and maybe if you can very quickly create a bunch of distilled models to test your application or to test part of your application ... that allows you to get to a high-value set of capabilities more quickly," Curran said.
Agentic coordination
AWS also said Nova Premier is suitable for multi-agent coordination use cases, even though it is not a reasoning model.
This point stresses that reasoning is not the only way to coordinate agents, said David Menninger, an analyst at ISG.
"Reasoning is one thing to help with multi-agent activity," Menninger said. "You still have to have a process by which you decide which agents to call." He added that one way to do this is with chain-of-thought processing, and another is traditional machine learning.
"We are very early in the multi-agent stage of agentic AI," Menninger continued. "I'm seeing mostly a single vendor with slightly different agents doing different things. I don't see anybody trying to coordinate and successfully coordinating Salesforce to ServiceNow to SAP. That type of process still is not yet, in my opinion, well handled by anyone in the market."
He added that another challenge for AWS is getting enough people comfortable using its Nova family of foundation models.
Meanwhile, AWS competitor Google Cloud expanded access to AI Mode, an experimental feature that enables users to ask complicated questions within Search. Google originally released the experimental feature in March.
Esther Shittu is an Informa TechTarget news writer and podcast host covering AI software and systems.
