Published on March 12, 2026
Reka has introduced a new version of its Edge model – a compact AI with advanced computer vision capabilities designed to run locally on devices without a cloud connection.
Products 5 – 7 minutes min readMost powerful AI models today operate remotely on servers in data centers that we access via the internet. This works well as long as there is a stable connection and latency isn't a dealbreaker. But what if you need the model right there on the device: in a robot, a drone, a security camera, or augmented reality glasses? That is where a completely different story begins.
Reka has released a new version of its Reka Edge model. Based on the description, it is an attempt to answer that exact question: how to make a serious AI compact and fast enough to function where the cloud is unavailable or its use is simply inefficient.
Why Do We Need «Edge» AI Anyway?
In the tech world, the term «edge» refers to edge computing – processing data directly on the end device or close to it, rather than in the cloud. Simply put: the calculations happen right where the result is needed, without sending data to a remote server.
This is important for several reasons. First, speed: sending a request to the cloud and waiting for a response takes time. For an autonomous car or a robot that needs to react in a split second, such a delay can be critical. Second, privacy: if a device processes images or video locally, the data never leaves its bounds. Third, cost: cloud computing incurs expenses that become quite significant at scale.
Reka Edge was built specifically for these scenarios, with a focus on «physical AI» tasks: robots, drones, smart cameras, and similar systems that must understand the surrounding world in real time.
What This Model Can Do
In essence, Reka Edge is a multimodal model that can «see»: it analyzes images and videos, recognizing and localizing objects within the frame. Furthermore, it is capable of interacting with tools and APIs, meaning it doesn't just describe what it sees but can also take actions based on that information.
Key areas where the model demonstrates high performance include:
According to the company, Reka Edge's results were close to those of the much larger Gemini 3 Pro model, despite its fundamentally smaller size.
Small but Nimble 🚀
The model size is 7 billion parameters. It's no giant: for comparison, flagship models boast hundreds of billions of parameters. However, compactness here is a deliberate choice rather than a limitation.
One of the key engineering ideas is that the model processes images significantly more efficiently than its counterparts. For a standard 1024×1024 pixel image, it requires approximately three times fewer tokens (internal processing units) than other models of a similar size. This directly impacts both speed and operating costs.
By the numbers: the model processes over 5 images per second in streaming mode – more than twice as fast as its closest competitors in its class. The time to first response is about half a second. For interactive applications where a user expects an instant reaction, this is a substantial metric.
Which Devices Can Run It?
Reka Edge is designed for a wide range of hardware: from servers and cloud infrastructure to specific hardware platforms – NVIDIA Jetson (a popular solution for robotics), Apple computers with Apple Silicon chips, standard Windows and Linux PCs, as well as smartphones and wearables based on Qualcomm Snapdragon.
It is also worth mentioning how it handles limited memory. In its standard form, the model takes up about 13 GB. By applying compression (quantization), this footprint is reduced to 5 GB – by nearly two-thirds – while maintaining over 98% of its performance quality. This opens up the possibility of running it on devices where memory is physically limited.
Where Could This Be Useful?
Reka highlights several application areas. First is physical AI: robots, drones, and automotive systems that need to understand their environment instantly. Second is media analytics: automatic video captioning, logo detection, and archive tagging. Third is augmented reality: smartphone and smart glass apps that require rapid environmental analysis. Fourth is automation based on visual input: agentic systems that read information from a screen or camera and perform actions in response.
To put it simply: the model is useful anywhere that requires a «see-understand-act» loop where the cloud is unavailable, too slow, or too expensive.
Reading Between the Lines
Reka honestly stipulates that comparisons with Gemini 3 Pro were conducted via API, while other models were run locally on a single cluster – meaning testing conditions were not identical. This is an important note to keep in mind when interpreting the figures.
Furthermore, the stated application scenarios (robots, drones, smart glasses) remain relatively niche fields. Real-world practice will show how much Reka Edge becomes an in-demand tool. The company has already opened access to the model through its own service, API, and an open repository, so it can be tested on real tasks right now.
In a broader sense, what is happening reflects a general trend: the industry is moving toward efficiency, not just scaling up. The question of «how to make a model more powerful» is being joined by «how to make it fit for real-world conditions», and Reka Edge is one of the most compelling answers to date.
You May Also Like
Events are only part of the bigger picture. These materials help you see more broadly: the context, the consequences, and the ideas behind the news.
AI: Events
Products
Liquid AI has introduced the LFM2-24B-A2B model, capable of running AI agents with tool-calling capabilities directly on consumer hardware – without the cloud or latency.
The latest update to AMD's Ryzen AI Software includes support for new models, improved performance, and expanded tools for developers working with AI on Ryzen processors.
Overworld has released Waypoint-1, a video generation model that runs locally and responds to controls in real-time during the content creation process.
How This Text Was Created
This material is not a direct retelling of the original publication. First, the news item itself was selected as an event important for understanding AI development. Then a processing framework was set: what needs clarification, what context to add, and where to place emphasis. This allowed us to turn a single announcement or update into a coherent and meaningful analysis.
We openly show which models were used at different stages of processing. Each performed its own role — analyzing the source, rewriting, fact-checking, and visual interpretation. This approach maintains transparency and clearly demonstrates how technologies participated in creating the material.
The neural network studies the original material and generates a coherent text
Correction of errors, inaccuracies, and ambiguous phrasing
Generating a textual prompt for the visual model
Generating an image based on the prepared prompt
Want to dive deeper into the world
of neuro-creativity?
Be the first to learn about new books, articles, and AI experiments
on our Telegram channel!
