Amazon introduced its Nova model family during its re:Invent 2024 conference in December. Then, it announced that the lineup would comprise:

  • Micro, a text-only model
  • Lite, Pro, and Premier: three foundation models capable of processing text, images, and video (no audio) as inputs to generate text outputs.
  • Amazon Nova Canvas, an image generation model; and
  • Amazon Nova Reel, a video generation model.

Out of these, only Premier was unavailable at launch, and Amazon committed to a Q1 2025. Delivering on its promise, Amazon announced the general availability of Amazon Nova Premier last Wednesday, praising the model as the most advanced in its Nova family of foundation models. According to the company, the Nova Premier, while not being a 'reasoning' model, is still ideal for complex tasks requiring capabilities including context understanding, multistep planning, and handling various tools and data sources to solve a single query. Nova Premier users can also benefit from the model's impressive one-million token context window, which allows it to process extremely long documents or large codebases.

Amazon tested Nova Premier across 17 different benchmarks to show that Premier performs better than Pro overall. Surprisingly, according to the published chart, Premier is actually slightly worse than Pro at following instructions (91.5% vs 92.1% on IFEval). Amazon also compared Premier with Claude 3.5 Sonnet, Claude 3.7 Sonnet (without thinking) and GPT-4.5 on the former's technical report, only to find quite unremarkable results. Some highlights of the comparison are that Premier achieves a better score than GPT-4.5 (but not any version of Sonnet) on the coding benchmark SWE-Bench Verified, or that it surpasses both Sonnet versions (but not GPT-4.5) on the Berkeley Function Calling Leaderboard.

Beyond its standalone capabilities, Nova Premier serves as a teacher model for distillation. This enables businesses to transfer Nova Premier's advanced capabilities into smaller, faster models like Nova Pro, Micro, and Lite for production deployment. Amazon cites an example in which it used Premier to distill Nova Pro for complex tool selection and API calling, achieving 20% higher accuracy while maintaining the speed and cost benefits of the smaller model.

Additionally, Amazon is also highlighting Premier's potential to serve as the supervising agent in multi-agent collaboration applications. Customers can also choose to combine Nova Premier's strengths and use Premier to distill a smaller and more cost-effective model to use it as a replacement for a supervisor agent in multi-agent systems that combines Premier's strong performance with the smaller model's speed and latency reduction. Amazon claims that customers who were granted early access are already seeing results. Slack reported that Nova Premier executes interactive analysis workflows while being faster and nearly half the cost of competing models. Robinhood Markets highlighted how Premier can unlock complex multi-agent systems which are performant, cost-effective, and fast.

Nova Premier is already accessible on Amazon Bedrock with on-demand pricing and select AWS regions availability via cross-region inference. The model includes built-in safety controls to promote responsible AI use, including content moderation capabilities that ensure the appropriateness of its outputs on a varied range of applications.