Amazon Nova: Foundation Models in Bedrock

What is Amazon Nova?

Amazon Nova is Amazon Web Services’ own foundation model family, announced at AWS re:Invent in December 2024. With Nova, Amazon competes directly with OpenAI, Anthropic, and Google in the generative AI space. All Nova models are exclusively available through Amazon Bedrock and integrate seamlessly into existing Bedrock workflows, Agents, and Knowledge Bases. The family covers a broad range of needs: from cost-efficient text processing to multimodal image and video analysis to image, video, and speech generation.

The family has grown significantly since launch. Nova became available in EU regions in February 2025, followed by the speech model Nova Sonic and the flagship Nova Premier in April 2025. At re:Invent 2025, AWS introduced the Nova 2 generation with built-in reasoning: Nova 2 Lite (generally available), Nova 2 Pro (preview), Nova 2 Sonic for voice conversations, and Nova 2 Omni as the first unified model that generates both text and images. The offering is rounded out by Nova Act for browser-based AI agents and Nova Forge for building your own models derived from the Nova models.

Core Features

• Understanding models in four sizes: Nova Micro (text-only, 128K tokens), Nova Lite and Nova Pro (multimodal, 300K tokens), and Nova Premier (multimodal, 1M tokens) cover every level of requirement.
• Nova 2 with reasoning: Extended thinking in three intensity levels (low, medium, high), a 1M-token context window, and built-in tools such as a code interpreter, web grounding, and remote MCP tools.
• Multimodal inputs: Nova Lite and Pro process text, images, and up to 30 minutes of video in a single request; Nova Premier analyzes documents over 400 pages and videos up to 90 minutes.
• Creative generation: Nova Canvas creates and edits images (inpainting, outpainting, background removal), while Nova Reel generates short videos with camera control.
• Real-time speech: Nova Sonic unifies speech understanding and speech generation in one speech-to-speech model with German support, function calling, and RAG grounding.
• Model distillation: Nova Premier serves as a teacher model to derive more compact, faster, and cheaper specialized models via Amazon Bedrock.

Typical Use Cases

• High-volume classification and extraction: Nova Micro delivers low-cost, low-latency responses for document classification, data extraction, and simple Q&A.
• Complex reasoning and agent workflows: Nova 2 Pro and Nova Premier suit multi-step tasks such as multi-document analysis, video reasoning, and software migrations.
• Multimodal analysis: Analyze images, screenshots, and videos for quality control, content moderation, or damage assessment.
• Marketing and e-commerce content: Image and video creation with Nova Canvas and Nova Reel for product visualization and campaigns.
• Voice-driven assistants: Real-time conversations with Nova Sonic in contact centers or in products, including German.
• Browser automation: With Nova Act, you can automate recurring tasks such as CRM updates or form submissions.

Benefits

• Attractive price-performance ratio: AWS positions the Nova models as a cost-efficient option within their respective intelligence classes, making them attractive for cost-optimized production applications.
• One API for all models: Through Amazon Bedrock you access Nova, Claude, Llama, and other models with the same interface, simplifying switching and comparison.
• Secure EU processing: Cross-region inference in Frankfurt, Ireland, Paris, and Stockholm keeps processing within the EU, and input data is not used for training.
• Scalable model tiers: From the low-cost Micro model to the Premier flagship, you choose the right cost-performance profile per use case.
• Deep Bedrock integration: Nova works directly with Bedrock Agents, Knowledge Bases, Guardrails, and model distillation.

Integration with innFactory

innFactory supports you in selecting the right Nova model for your use case and integrating it into your existing Bedrock infrastructure. As an AWS Reseller, we advise you on cost-optimized usage, EU-compliant architecture via cross-region inference, and evaluation against third-party models such as Claude or Llama. We also help you build agent workflows, RAG integrations, and voice conversations, as well as set up model distillation to turn high-performing models into cheaper specialized ones.