Reka Edge

Reka Edge is a vision language model designed by Reka for physical AI applications running on edge devices, where it serves as a lightweight alternative to larger cloud-based vision models. It is positioned for practical deployment scenarios that demand deep visual reasoning and physical grounding, with strengths in image understanding, video analysis, object detection, and agentic tool use. The model is offered with open weights so it can be run locally through common inference stacks such as Hugging Face and vLLM, or accessed via the Reka API.

The architecture pairs a 657M parameter ConvNeXt V2 convolutional vision encoder, which is tuned for efficient processing of streaming video, with a 6.4B parameter transformer backbone trained from scratch to handle reasoning and generation. To keep inference fast and the context window from filling up too quickly, Reka Edge produces only 64 tokens per image tile, a deliberate token efficiency choice that helps it handle high-definition visual data with shorter response times. Reka trained the model on petabytes of multimodal data that blends synthetic and real sources, aiming to deliver frontier-level visual intelligence in a compact 7B package.