Reuna runs standard AI models directly on the chip. No operating system, no cloud, no GPU. From a $5 microcontroller to our own re-programmable FPGA accelerator.
The Reuna FPGA accelerator runs your ONNX model as bare-metal logic. No operating system, no drivers, and because the fabric is re-programmable, the silicon itself can change after the product ships.
The only real option is a GPU board: an operating system, drivers, CUDA, Python, and 10 to 15 watts. Most devices are battery-powered, cheap, small, and offline. Drag the slider and watch what the same job costs on each.
If your team can export a model, they can ship it on Reuna. The compiler turns a standard ONNX file into lean bare-metal code for our microcontroller or our FPGA-based accelerator, and either one talks to the rest of your board over a four-wire SPI link that nearly every device already has.
No runtime to port, no drivers to write. The compiler reads the ONNX graph, quantises it, schedules it for the target, and emits bare-metal C you flash like any other firmware.
PyTorch, TensorFlow, whatever the team already uses.
The industry-standard model file. One click in most tools.
Model in, bare-metal code out. No OS in between.
No hardware yet? Use our microcontroller board or our FPGA-based accelerator for real-time work. Both run offline and plug in over SPI.
Your model. Your chip. Our compiler. Femto compiles for the chip you already ship. Upload a model, download firmware, flash. Try the fit check below: estimates are scaled from our YOLOv5s build, real numbers come from Femto.
Two different platforms for two different jobs. Our microcontroller board is for edge devices that check in every few seconds. Our FPGA-based accelerator is for machines that need answers in milliseconds.
Compile your AI onto the microcontroller already in your product, or use the Reuna module. We run YOLOv5 on a $5 ESP32-S3: 23 seconds per image stock, about 5 after Reuna optimisation, on identical hardware. Update the model with a firmware push, never a hardware swap. Updatable, not re-programmable: that's the accelerator's job.
Our FPGA-based accelerator is for robots, vehicles and any device that needs quick decisions on board. It runs AI at a quarter of the power of a GPU board, at a fraction of the cost. And because it's re-programmable, you can change what it does after the product ships. Not just the model, but the logic running on the chip.
Cabin cameras that wake a drowsy driver. Startups are building this on ESP32 today; we make the AI actually fit.
Livestock counting and fence-line sensors in fields with no signal.
Defect detection and safety-zone monitoring on the machine itself.
Perimeter cameras that raise a real alert instead of recording hours of nothing.
Perception in milliseconds, offline.
Toys, appliances, wearables. Voice and vision on a few-dollar chip.
Same stack, restricted deployments. Offline, deterministic, no OS to attack.
Building a robot, a camera, a sensor, a toy? If it exports to ONNX, we can run it on your hardware or ours.