Reka Edge Guide: A 7B Vision Model Built for Devices

Most vision-language models are judged on how well they answer. Reka Edge is pitched on how cheaply it sees. Reka announced it on 11 March 2026 as a model for robots, drones, cameras and smart glasses, where the constraint is not a benchmark score but a power budget, a memory ceiling and a frame that needs an answer before the next one arrives.

That framing is useful, because it tells you exactly which questions to ask. How many tokens does an image cost? How fast is the first token? Does it fit on the hardware you have? And, less obviously, are you allowed to ship it?

What Reka built

Reka Edge is about 7B parameters in total. Reka's launch post breaks that into a 657M-parameter ConvNeXt V2 vision encoder and a 6.4B-parameter transformer backbone trained from scratch. It accepts images, video and text, and returns text. It does not have a reasoning mode.

The architectural detail that matters is the image encoding: Reka says the model produces 64 tokens per image tile. Fewer visual tokens per image means less compute per frame, a shorter prompt, and a faster first token. Everything else in the pitch follows from that choice.

Token efficiency is the headline

Reka's model card measures a single 1024 by 1024 image. Reka Edge used 331 input tokens, against 1,063 for Cosmos-Reason2 8B and 1,041 for Qwen3.5 9B. Time to first token was 0.522 seconds, against 0.844 and 0.60 respectively, all measured on local inference.

The launch post adds a throughput figure: 5.46 images per second at 32 concurrent requests, which Reka describes as more than twice the rate of those two comparison models. For a multi-camera system or a video stream sampled at a few frames per second, that is the number that decides how many streams one device can handle.

Roughly a third of the visual tokens also means roughly a third of the image portion of the input bill on a hosted endpoint, and a shorter prompt leaves more room in a small context window. That last point matters here, as covered below. For why image token counts vary so much between models, see multimodal models explained.

Reading the benchmark table carefully

Reka publishes a comparison against Cosmos-Reason2 8B, Qwen3.5 9B and Gemini 3 Pro. Some selected rows:

• MLVU (video): 74.30, against 52.39 for Qwen3.5 9B and 80.68 for Gemini 3 Pro.
• MMVU (video): 71.68, against 68.64 and 78.88.
• VQA-v2: 88.40, against 83.22 and 89.78.
• RefCOCO-A (detection): 93.13, against 93.62 for Qwen3.5 9B.
• Mobile Actions (tool use): 88.40, against 91.78 for Qwen3.5 9B.

The pattern is honest enough to be useful. Video understanding is where Reka Edge pulls clearly ahead of its size class. On grounding and tool use, Qwen3.5 9B edges it. Gemini 3 Pro leads on the video and VQA rows, though it was measured through an API rather than locally. These are Reka's own numbers, so run your own frames before trusting any of them.

Running it on your own hardware

The weights are on Hugging Face as RekaAI/reka-edge-2603. Reka's launch post says 4-bit quantisation reduces memory from 13GB to 5GB while keeping over 98 percent of original performance. Reka's own quantised builds are offered on request; the Hugging Face page also points to GGUF quantisations for llama.cpp-based runtimes, which are worth testing separately.

The model card lists Jetson Thor and Jetson AGX Orin as supported, along with Apple Silicon Macs with at least 24GB of memory and Linux machines with a 24GB GPU. With quantisation it names smaller targets including Jetson Orin Nano and recent Snapdragon phones. It runs through Transformers, vLLM with a Reka plugin, llama.cpp, SGLang and Docker Model Runner.

Two practical warnings from the card: the Transformers path is pinned to a specific version and needs trust_remote_code=True, and on Apple MPS you should use float16 rather than bfloat16. Quantization explained covers what you trade when you go to 4 bits.

The licence is the part to read first

Reka Edge is open weights, not open source. It ships under a Business Source License 1.1 variant with these terms:

• Research, evaluation and non-production use are free.
• Commercial production use is free only for for-profit companies under $1 million in annual gross revenue, counting affiliates.
• Above that, and for government and academic production use, a commercial licence from Reka is required.
• Hosting it as an API for third parties, or using its outputs to train competing commercial models, needs a commercial licence regardless.
• Two years after public release the licence converts to Apache 2.0.

For a startup prototyping a camera product, that is generous. For an established company planning to ship devices, it is a procurement conversation before it is an engineering one. Open weights vs open source explains why that distinction tends to surface exactly when a product succeeds.

Using the hosted version

Reka announced availability on OpenRouter in April 2026. OpenRouter lists it at $0.10 per million tokens for both input and output, with tool calling and structured outputs, at the time of writing. It also lists a 16K context window, which is small: fine for a handful of frames and a question, not for long video or long conversations. Sample frames rather than sending a clip wholesale.

Hosted access is the right way to evaluate it before committing hardware. Build a set of fifty frames from your real cameras, write the questions you actually need answered, and score Reka Edge against whichever vision model you would otherwise use. If it holds up at a third of the visual tokens, the device story is worth pursuing.