Reka Edge

Designed for deployment on edge devices, the platform combines a 660 M‑parameter ConvNeXT V2 vision encoder with a 6 B‑parameter language backbone, totaling 7 B parameters, to deliver fast multimodal processing while keeping resource use modest. It ingests high‑frame‑rate video streams from onboard cameras, performs object detection, scene understanding, and generates precise bounding boxes in real time, enabling robots, drones, and other physical systems to locate tools, assess environments, and execute actions with sub‑second latency.

The system targets developers building autonomous robotics, automotive, wearables, public‑safety drones, and media‑content applications that require continuous visual perception without reliance on cloud services. By running locally on hardware such as NVIDIA Jetson, it avoids the delays and safety risks associated with remote large‑language‑model calls, providing spatial intelligence and conversational grounding for tasks like “where is the 10 mm wrench?”

Open‑source weights are hosted on Hugging Face, and the model can be accessed either through a local runtime or an API, facilitating integration into production pipelines that demand high‑speed, reliable edge AI.