Amazon Nova Models – Speed & Performance Insights

Amazon has introduced the Nova family of foundation models on Amazon Bedrock, tailored for various use cases across cost, speed, and multimodal capabilities. Here’s a breakdown of the core Nova models:

• Nova Micro: Text-only model offering lowest latency and cost.
• Nova Lite: Ultra-fast multimodal model handling text, image, and video inputs at low cost.
• Nova Pro: Balanced, high-performing multimodal model optimized for speed, cost, and accuracy.
• Nova Canvas: Advanced model for image generation.
• Nova Reel: Specialized model for video generation.

---

⏱️ Speed Comparison

Time Comparisons for different models were tested with a prompt containing chat history, query and knowledge base retrieved data.

Model Name	Run 1 (s)	Run 2 (s)	Run 3 (s)	Average Time (s)
anthropic.claude-3-sonnet-20240229-v1:0	64.82	58.3	58.91	60.68
anthropic.claude-3.5-sonnet-20240620-v1:0	56.67	57.94	59.79	58.13
anthropic.claude-3-haiku-20240307-v1:0	41.27	41.95	44.17	42.46
anthropic.claude-3.5-haiku-20241022-v1:0	53.66	51.91	54.18	53.25
amazon.nova-micro-v1:0	33.59	35.3	34.28	34.39
amazon.nova-lite-v1:0	34.83	35.12	34.14	34.70
amazon.nova-pro-v1:0	36.95	37.44	36.5	36.96

---

🔍 Key Observations

• Nova Micro is the fastest among all tested models, followed closely by Nova Lite.
• Nova Pro offers the best quality-to-speed balance, making it ideal for production-grade multimodal tasks.
• Among Anthropic models showed the most improved quality, but still lags behind Nova models in speed.
• Among the Anthropic models, Sonnet 3 and Sonnet 3.5 takes largest time.
• Nova Pro and Haiku 3.5 provide better response quality compared to Nova Micro and Lite, useful for tasks needing deeper reasoning.

---

🔗 References

• Amazon Bedrock – Nova Foundation Models
• Amazon Nova User Guide (Docs)
• Nova Technical Report and Model Card (PDF)

---