Build AI your way with Amazon Nova

AI delivers the most value when it deeply understands your business. Whether you're enhancing the customer experience of your AI applications or building agents that can handle diverse tasks for your business, Amazon Nova gives you the tools to adapt the model to your needs. You can now combine the power of your organization’s data with the foundations of Amazon Nova using a comprehensive suite of customization tools and capabilities - from aligning model responses to your preferred tone and behavior, to training it on your unique workflows and business knowledge. With flexible options across the model training lifecycle, you can build AI that thinks and acts like your business.

Learn more

"Amazon Nova’s customization capabilities have unlocked major advances in our materials science research by making exerpiments that required in-depth knowledge accessible through LLMs that can be prompted. By applying Supervised Fine-Tuning, we boosted our metamaterial design AGI’s valid material generation rate from 2.7% to 97.5% for materials design and from 19.3% to 98% for reconstruction tasks. Analysis error rates dropped by up to 95%, enabling the creation of an expert assistant that generates and analyzes cellular metamaterial designs using material properties, images, and structured code that is accessible to both humans and large language models. This enables the researchers to be more efficient with their time, and the natural language interface lowers the barrier to entry for others. Encouraged by these transformative results, we’re expanding our use of Nova to push the boundaries of AI in specialized scientific domains."

Wojciech Matusik, Cadence Design Systems Professor, MIT

"Our fine-tuned Nova Pro, trained using SageMaker Training Jobs, outperformed all other models, on the SWE-Lancer benchmark, that we currently serve to our enterprise customers on AWS.  This rigorous benchmark translates coding performance into real-world earning potential—representing how much a model could earn if it were doing software engineering freelancing jobs or bounties. Combined with Nova’s industry-leading price-performance, this accuracy gain translates into record levels of ROI for tackling compound coding tasks within complex legacy codebases. We are looking forward to serving our AWS customers with this fine-tuned model, and drive closer to the human baseline on SWE-Lancer with our deep collaboration with Amazon Nova on AWS"

Yang Li, Cofounder, Cosine

"As we explore ways to automatically evaluate brand consistency and proper localization across our global marketing content, we tried the Supervised Fine-Tuning (SFT) recipe for aligning Nova Pro with our marketing experts’ knowledge. We achieved a 15% improvement (55% to 70%) in the model’s ability to identify on-brand images, demonstrating stronger alignment with Volkswagen’s brand guidelines. We are now looking forward to building on these results with Amazon Nova as it enables Volkswagen Group’s vision to scale high-quality, brand-compliant content creation across our diverse automotive markets worldwide through GenAI."

Dr. Philip Trempler, Technical Lead AI & Cloud Engineering, Volkswagen Group Services

"Using PEFT and full fine-tuning, we significantly improved our product listing validation accuracy — achieving an 18-point increase in detection recall and better precision. These significant accuracy gains, combined with Nova's cost-efficiency, can transform how we manage product catalogs."

Umit Batur, Sr. Manager of Applied Science, Amazon Catalog System Services

"We are customizing Amazon Nova to power an AI-driven issue resolver, with the goal of automating our customer service interactions. Today, issue resolution is a two-step process of intent detection followed by issue resolution for the detected intent that involves multi turn tool calling and response generation. Through SFT our custom Nova Micro for high-volume and low-latency intent detection demonstrated 76.9% in-domain accuracy and 69.2% on the generalization test, outperforming our current baseline by an absolute 5.4% and 7.3% respectively on offline evaluations. For tool selection, our custom Nova Lite achieved 86.1% accuracy, a 4.8% absolute improvement over our current baseline. We are looking forward to dialing up the A/B testing with these models for online testing on real production traffic, in addition to continued development towards a unified multilingual model capable of handling most of customer service tasks through additional approaches such as Continued Pre-Training."

Jing Huang, Applied Scientist, Amazon Customer Service