Amazon: Nova Premier 1.0

Processes both text and image inputs simultaneously to perform complex reasoning tasks, using a unified transformer architecture that encodes visual and textual tokens into a shared embedding space. The model applies attention mechanisms across modalities to establish cross-modal relationships, enabling it to answer questions about images, perform visual analysis, and reason about relationships between visual and textual concepts in a single forward pass.

Serves as a teacher model for knowledge distillation workflows, where its internal representations and outputs are used to train smaller, task-specific student models. The model exposes logits, attention patterns, and intermediate layer activations that can be extracted and used to guide the training of custom models through techniques like response-based distillation (matching output distributions) and feature-based distillation (matching hidden layer representations).

Processes extended text inputs (documents, code files, conversation histories) with maintained coherence across thousands of tokens, using an efficient attention mechanism (likely sparse or hierarchical attention) that reduces computational complexity while preserving long-range dependencies. The model maintains semantic understanding across document boundaries and can perform tasks like summarization, question-answering, and analysis that require understanding relationships between distant parts of the input.

Generates text outputs constrained to match a provided JSON schema or structured format specification, using guided decoding or constrained beam search that enforces token-level validity against the schema. The model's output is guaranteed to be parseable as valid JSON or structured data matching the schema, with type validation (strings, numbers, arrays, objects) enforced at generation time rather than post-processing.

Generates syntactically correct and logically sound code across multiple programming languages, using patterns learned from large code corpora to produce implementations that follow language idioms and best practices. The model understands code structure, dependencies, and common algorithms, enabling it to generate complete functions, classes, or multi-file solutions from natural language specifications or partial code contexts.

Breaks down complex problems into logical sub-steps and generates detailed reasoning chains, using chain-of-thought prompting patterns to expose intermediate reasoning before arriving at conclusions. The model articulates its reasoning process, identifies dependencies between steps, and can backtrack or revise reasoning when contradictions are detected, enabling more reliable solutions to multi-step problems.

Provides access to Nova Premier through standardized API endpoints via OpenRouter or AWS Bedrock, abstracting underlying infrastructure and enabling seamless switching between providers or model versions. The API handles request routing, load balancing, and response formatting, with support for streaming responses, batch processing, and standard parameters (temperature, top-p, max-tokens) that work consistently across providers.