The Pulse
Reka Edge Supports Tool Calls in a 7B Vision Model
Reka Edge is a 7B vision-language model that accepts images and video and supports tool calling through Reka’s API. The company announced the model on March 11, 2026, and its API documentation lists a 16,384-token context window and support
AI.info Team · Written with AI ·
Reka Edge puts tool support in its API catalog
Reka’s API documentation lists Reka Edge 2603 as a ready model that accepts text, images and video, and supports tool calls, structured outputs and streaming. The entry shows a 16,384-token context window and the same maximum output length. That makes the 7B vision-language model available not only for describing visual inputs, but also for requests where an application can offer tools for the model to call.
The model’s release is not new: Reka announced Reka Edge on March 11, 2026. The company described it as “This extremely efficient 7B parameter model is optimized to deliver industry-leading performance across key capabilities, including image understanding, video analysis, object detection, and agentic tool-use.” Reka, in its March 11 announcement
In practical terms, tool support lets a developer provide callable functions or services in a request and have the model return a tool call when it judges one appropriate. The application still has to execute that call and return its result; the model listing does not mean Reka Edge independently controls devices or services.
A 657-million-parameter vision encoder handles images
Reka says the model pairs a 657-million-parameter ConvNeXt V2 vision encoder with a 6.4-billion-parameter transformer trained for reasoning and generation. The company designed the vision system to produce 64 tokens per image tile, a compact representation intended to reduce image-processing overhead. Reka’s announcement says a 1,024-by-1,024 image uses about three times fewer tokens than in similarly sized models.
The API listing gives the model a 16,384-token context, while its inputs can include video as well as still images and text. That mix positions Edge for applications such as visual inspection, video analysis and systems that need to choose an action based on what a camera sees. Its outputs are text, according to the model entry; the documentation does not list image or video generation.
Reka’s speed figures come from its own tests
Reka reports throughput of 5.46 images per second and a 0.522-second time to first token in its evaluation. For the speed comparison, the company says it hosted the open-weight models on its inference cluster using vLLM, while Gemini 3 Pro was measured through its API. Those figures are company-reported results, not an independent evaluation, and the different serving arrangements matter when comparing latency.
The company also says its 4-bit quantized version reduces memory use from 13GB to 5GB and retains more than 98% of the model’s original performance. The announcement describes the smaller memory footprint as a route to running the model on devices with tighter memory and power limits. Reka separately says its 3.5-bit quantization option requires contacting the company for access.
Local weights and API access come with different terms
Developers can use Reka Edge through the Reka API or run its weights themselves, including with vLLM, according to the company’s model card on Hugging Face. The card says commercial use is permitted for organizations earning less than $1 million a year, a condition that limits who can use the weights commercially without further permission.
For API use, Reka’s model-list response shows a sample rate of $0.0000001 per input token and the same per output token—equivalent to $0.10 per million tokens. The company says rates can change and directs developers to its model catalog for current prices. Its model documentation lists Edge as supporting tools and structured outputs, distinguishing it from Reka Flash 3, which the same catalog lists as not supporting tools.
Reka’s announcement frames Edge for on-device, on-premises and cloud deployment, but the choice between the API and local hosting carries different trade-offs in cost, control and licensing. The published API entry confirms tool support; the model’s benchmark claims and deployment performance remain the company’s own measurements.