Amazon Releases Foundation Model Nova

After observing major tech players’ AI strategies throughout the year, Amazon’s approach as the cloud service leader has been particularly intriguing. Amazon unveiled its foundation models, Amazon Nova, at AWS re:Invent 2024 held during Dec. 2-6, 2024.

AWS re:Invent 2024 – Generative AI Strategy

This year’s keynote consistently revolved around one core message: AWS’s identity as a Cloud Service Provider (CSP) with a fundamental mission – and greatest value proposition to customers – centered on reliability, security, high performance, efficiency, and scalability. This philosophy is reflected in their AI strategy messaging, which emphasizes “what AWS can help you solve” rather than “what AWS has built.”

The cornerstone of AWS’s generative AI services is the Bedrock platform, primarily offering AI model hosting that enables developers and enterprises to leverage various models via API based on their specific needs. This year, AWS announced several significant updates to Bedrock that are crucial for enterprises looking to implement generative AI.

First is the introduction of model distillation support, which allows extracting smaller, task-specific models from larger ones to achieve faster performance and cost optimization.

Second, Bedrock, which already supported Retrieval-Augmented Generation (RAG) – allowing AI models to access specified knowledge bases for response generation without retraining – will now support an expanded range of vector databases.

Third, addressing one of generative AI’s most criticized issues – “hallucination” – Bedrock will introduce automated reasoning capabilities, essentially performing deeper analysis and verifying user prompts.

Fourth, with “agents” or “agentic” buzzwords of the past six months, reflecting our desire to move beyond one-to-one AI interactions toward one-to-many or even autonomous operations, Bedrock will now facilitate task coordination among multiple AI agents.

In essence, AWS’s value proposition extends beyond offering superior, cost-effective computing power or flexible, rapid storage and database capabilities – it’s about enabling customers to deploy generative AI solutions tailored to their specific requirements.

Amazon Nova

Interestingly, this year’s AWS re:Invent featured a “one more thing” moment from Amazon CEO Andy Jassy: Amazon Nova.

Amazon Nova represents next-generation foundational models, comprising six distinct models:

• Amazon Nova Micro: a text-only model that delivers the lowest latency responses at very low cost.
• Amazon Nova Lite: a very low-cost multimodal model that is lightning fast for processing image, video, and text inputs.
• Amazon Nova Pro: a highly capable multimodal model with the best combination of accuracy, speed, and cost for a wide range of tasks.
• Amazon Nova Premier: the most capable of Amazon’s multimodal models for complex reasoning tasks and for use as the best teacher for distilling custom models (available in the Q1 2025 timeframe).
• Amazon Nova Canvas: a state-of-the-art image generation model.
• Amazon Nova Reel: a state-of-the-art video generation model.

The family of Micro, Lite, Pro, and Premier models is particularly noteworthy, differentiated primarily by modality and parameter count. For instance, Micro handles text-only tasks, while Lite, Pro, and Premier can process text, audio, image, and video inputs.

The official press release repeatedly emphasizes one key point: these four models not only deliver rapid performance matching or exceed competitors in their respective classes but do so at a significant cost advantage. While specific token pricing isn’t yet available, the company claims a 75% cost reduction compared to “models with comparable intelligence capabilities.”

Another highlight is the extensive context window support across these models, with even the smallest model, Micro, supporting 128K tokens, and Lite and Pro, supporting 300K. Amazon has announced a target of 2M token support by 2025.

Two noteworthy announcements from Jassy’s presentation include the Q1 2025 launch of “voice-to-voice” functionality enabling seamless AI conversations and, more remarkably, the 2025 introduction of “Any-to-Any” capabilities – enabling various combinations like “text-to-image,” “image-to-audio,” and “audio-to-video,” analogous to interconnected brain regions.

Beyond these four models, Amazon Nova includes two visual-focused models: Canvas for image generation and Reel for video generation. Unlike the previous family’s emphasis on price-performance ratio, these models prioritize “high quality” and “controllability,” including features for local color and composition control and prompt-based pacing and camera movement control.

Conclusion

AWS’s core mission remains unchanged: providing developers and enterprises with the most competitive tools to build market-leading products and services. This commitment persists in the generative AI era, with only the challenges and customer needs evolving.

Simultaneously, Anthropic announced a deeper collaboration with AWS under “Project Rainier” to explore next-generation AI computing chips and leverage them for future model training.