Reka Edge

Accepts static images as input alongside text prompts and generates natural language descriptions, answers, or analysis. The model processes visual features through a vision encoder that extracts spatial and semantic information, then fuses this with text embeddings in a shared latent space before decoding text output. This enables tasks like image captioning, visual question answering, and scene understanding without separate image-to-text pipelines.

Processes video inputs by sampling key frames and maintaining temporal coherence across the sequence, allowing the model to understand motion, scene changes, and temporal relationships. The architecture extracts visual features from multiple frames and encodes temporal ordering information, enabling the model to answer questions about video content, summarize events, or track objects across time without requiring external video processing libraries.

Extracts text from images while maintaining spatial relationships and document structure, using the vision encoder to identify text regions and the language model to decode content while preserving layout information. This enables structured extraction from documents, forms, and screenshots without separate OCR engines, and the model understands context to correct misrecognitions based on semantic meaning.

Accepts an image and a natural language question, then generates an answer by reasoning about visual content. The model uses the vision encoder to extract relevant visual features, attends to regions of interest based on the question, and generates a response that demonstrates understanding of spatial relationships, object properties, and scene context. This enables open-ended visual reasoning without predefined answer categories.

Exposes image understanding capabilities through a stateless REST API that accepts HTTP requests with image payloads and returns JSON responses, enabling integration into batch processing pipelines, serverless functions, and distributed workflows. The API handles image encoding, model inference, and response serialization transparently, with support for concurrent requests and standard HTTP semantics (retries, timeouts, rate limiting).

The 7B parameter architecture is specifically optimized for inference speed through quantization, knowledge distillation, and efficient attention mechanisms, delivering sub-second response times on standard hardware. The model uses techniques like grouped query attention and optimized matrix operations to reduce computational overhead while maintaining accuracy, enabling real-time applications and high-throughput batch processing without requiring high-end GPUs.