Amazon: Nova Micro 1.0

Amazon Nova Micro uses a lightweight model architecture optimized for minimal inference latency through quantization, pruning, and edge-compatible parameter reduction. The model is designed to generate text responses with sub-second latency by reducing model size while maintaining semantic coherence, enabling real-time conversational interactions without sacrificing response quality for simple tasks.

Nova Micro is exposed through a pay-per-token API model via Amazon Bedrock or OpenRouter, allowing developers to invoke the model without managing infrastructure, with pricing scaled to the model's reduced parameter count. The API handles request routing, load balancing, and token accounting transparently, enabling predictable cost scaling based on actual usage rather than reserved capacity.

Nova Micro maintains conversational context through a fixed-size context window that accumulates conversation history, system prompts, and user messages. The model processes the entire context window as input for each generation, enabling coherent multi-turn conversations while requiring developers to implement context management strategies (truncation, summarization, or sliding windows) to stay within token limits.

Nova Micro supports streaming responses where tokens are emitted incrementally as they are generated, allowing clients to display partial results in real-time rather than waiting for complete response generation. The streaming API uses server-sent events (SSE) or similar protocols to push tokens to the client, enabling progressive rendering and perceived latency reduction in user interfaces.

Nova Micro is trained on multilingual data and uses a language-agnostic tokenizer that handles text in multiple languages without requiring language-specific preprocessing. The model can generate coherent responses in dozens of languages, with performance varying based on training data representation for each language, enabling developers to build globally-accessible applications without language-specific model variants.

Nova Micro accepts system prompts that define behavioral constraints, role-play scenarios, output formats, and reasoning approaches. The system prompt is prepended to the conversation context and influences all subsequent generations within that conversation, enabling developers to customize model behavior without fine-tuning. This is implemented through prompt engineering patterns rather than architectural modifications to the model.

Nova Micro can perform text classification and sentiment analysis by formulating classification tasks as natural language prompts, without requiring labeled training data or fine-tuning. The model generates text responses that indicate classification results (e.g., 'positive', 'negative', 'neutral'), leveraging its language understanding to infer categories from task descriptions. This approach is implemented through prompt engineering rather than specialized classification layers.

Nova Micro can generate abstractive summaries of longer text by processing the full text as input and generating a condensed version that captures key information. Unlike extractive summarization (selecting existing sentences), abstractive summarization generates new text that paraphrases and condenses the original, implemented through the model's language generation capability without specialized summarization layers.