Reka AI

What Reka is

Reka AI describes itself on its site as an AI research lab building models and infrastructure for the physical AI era^[1]. The direction it states is toward omni models: single architectures that natively process and generate every kind of data, and that run in real time^[1]. Two public releases show how that programme has been carried out so far — a reasoning model published in March 2025, and a small vision-language model announced in March 2026 and aimed at devices rather than at cloud inference.

reka-flash-3: a reasoning model trained from scratch

On 11 March 2025 Reka released reka-flash-3, a 21-billion-parameter reasoning model^[3]. It was trained from scratch rather than adapted from an existing base model, and it was published under the Apache 2.0 licence^[3]. Reka says the model performs competitively with OpenAI's o1-mini^[3]. The same model is described as powering Nexus, Reka's platform for building and managing AI workers^[3]. A reasoning model in this parameter range is something a company can serve to many users at once, which fits a platform product where reasoning is the thing being sold.

Reka Edge: perception on the device

Reka announced Reka Edge on 11 March 2026 and positioned it for physical AI on the edge rather than for cloud inference^[2]. It is a 7-billion-parameter vision-language model that takes images, video and text as input and returns text^[2].

The architecture divides the work between two parts. The vision encoder is a 657-million-parameter ConvNeXt V2, paired with a 6.4-billion-parameter transformer trained from scratch^[2]. Nearly all of the parameter budget therefore sits on the transformer side, while the part that turns pixels into features stays comparatively small — the kind of split that matters when a model has to fit into the memory of a device sitting next to a camera.

On speed, Reka reports that Reka Edge processes 5.46 images per second, more than twice the rate of Cosmos Reason2 8B and Qwen3.5 9B^[2]. For a model reading a video stream, frames per second is a practical limit: a system that cannot keep up with the camera has to drop frames or fall behind, which is why the figure is reported alongside the model's size and architecture.

The weights are open, and the licence allows commercial use below 1 million US dollars of annual revenue^[4]. That arrangement puts the model in the hands of small teams and researchers while leaving larger deployments to a separate agreement.

Reka names robots, drones, cars, smart glasses and media tagging as targets for the model^[2]. In each of these settings the camera and the model are in the same place; sending every frame to a data centre would add delay and would stop working when the connection does.

Where it sits

The two releases show a shift in what the same lab is optimising for. reka-flash-3 is a reasoning model whose job is to think through problems and whose home is a platform for AI workers^[3]. Reka Edge is a perception model whose job is to look at the world continuously and answer quickly, on hardware that is not a data centre^[2]. Reka's own framing holds them together: omni models that process and generate every kind of data in one architecture and run in real time^[1]. That stated goal is broader than either of the published models, and the two releases are the pieces of it Reka has shown so far. Reka Edge, in particular, makes the edge case concrete by pairing a small, fast vision stack with an open licence that lets outside teams try it on their own hardware^[2][4].