Reka Edge

What it is

Reka Edge is a vision-language model released on 11 March 2026 ^[1][2]. It has 7 billion parameters and accepts images, video and text as input, returning text ^[2]. Pixels in, words out: an image encoder turns a picture or a video frame into a sequence of vectors, and a language model reads that sequence together with any written question and continues it as prose. The same weights can describe a scene, answer a question about a photograph or point at an object, without a separate model for each job.

How it is built

The parameters are divided between two components. The vision encoder is a 657-million-parameter ConvNeXt V2, a convolutional network rather than a transformer ^[2]. The language side is a 6.4-billion-parameter transformer trained from scratch ^[2]. Together they make up the 7 billion parameters in the model's name.

The design decision that shapes much of the rest is how many tokens an image is allowed to cost. Reka Edge emits 64 tokens per image tile, roughly three times fewer than comparable models ^[2]. This matters because the cost of running a transformer grows with the length of the sequence it is handed: every extra token adds work in the attention layers and occupies memory that a small device may not have. A model that spends three times fewer tokens on the same picture spends correspondingly less compute and less memory on it, and on phones and single-board computers it is often memory movement rather than arithmetic that sets the limit. Shrinking the visual side is therefore one of the few levers that improves speed and footprint at the same time.

Speed

Reka reports a throughput of 5.46 images per second, more than twice the rate of Cosmos Reason2 8B and Qwen3.5 9B ^[2]. It also reports a time to first token of 0.522 seconds ^[2]. That second figure is the delay between sending a request and seeing the first piece of the answer, the number a user actually feels: a model may finish a long answer quickly overall and still feel sluggish if nothing appears for several seconds. For a camera or video feed, where frames arrive continuously, both numbers decide whether the model can keep up with its input at all.

Benchmark results

On the visual question-answering benchmark VQA-v2 the model scores 88.40; on MLVU it scores 74.30 and on MMVU 71.68 ^[1]. On referring-expression detection, where the model must find the object a phrase describes, it scores 93.13 on RefCOCO-A and 86.70 on RefCOCO-B ^[1]. Scores of this kind are only comparable within their own benchmark, and they say nothing about how the model behaves on the long tail of real photographs; but they establish that a 7-billion-parameter model of this size is being measured on the same tasks as larger systems rather than on a reduced set.

Running it locally

Reka Edge is meant to be run on the machine that produces the data. It runs on NVIDIA Jetson, Apple silicon, Qualcomm Snapdragon and x86 machines ^[2]. To fit those targets, the weights can be compressed: 4-bit quantization cuts memory from 13 GB to 5 GB, a 62 percent reduction, while keeping over 98 percent of the original performance ^[2]. Quantization stores each weight with fewer bits, trading a little numerical precision for a much smaller footprint, which is what makes the difference between a model that needs a discrete graphics card and one that fits on a device the size of a credit card.

Licence

The weights are open, under a licence that allows commercial use below 1 million US dollars of annual revenue ^[1]. Arrangements of this shape are a middle path between fully permissive release and an API: small teams and individual developers can build and ship with the model, while larger organisations are expected to come to a separate agreement.

Locating objects

The model can be asked to find things with a "Detect: expression" prompt, and it returns labelled bounding boxes in a bbox block ^[2]. The output is plain text that a program can parse, so a detection result can be fed straight into the rest of an application without a special decoder.

Where it sits

Reka Edge belongs to a group of models defined less by a benchmark ceiling than by where they can run. The trade is explicit: a 7-billion-parameter model with a deliberately compact visual encoder gives up some headroom in exchange for latency, memory and the ability to work offline on hardware that is already deployed. Its use of a ConvNeXt V2 encoder rather than a transformer, its 64-token image budget and its quantized 5 GB footprint are all choices in service of the same goal — keeping a vision-language model inside the device that sees the world.