Mobilint puts Ceva’s AI processors into production in its REGULUS chip

Ceva’s NeuPro AI processor family has reached its first production deployment, the company said in a November 2, 2026 announcement. Mobilint has integrated the technology into REGULUS, a chip designed to run AI inside robots, cameras and other devices with tight power limits. A REGULUS-powered USB accelerator is already shipping, according to the companies.

The deployment brings two Ceva designs into Mobilint’s processing architecture: SensPro, a programmable digital signal processor for sensing work, and NeuPro-Nano NPN32, a neural processing unit for running AI models. REGULUS combines them with Mobilint’s own AI processing technology rather than relying on a single engine for every workload.

Ceva’s AI division general manager, Yaron Galitzky, described the production milestone as progress in bringing the company’s AI technology into commercial silicon. Mobilint CEO Dongjoo Shin emphasized a different part of the same design: fitting sensing, signal processing and AI workloads into devices with very limited power budgets.

Different engines for sensing and inference

REGULUS assigns sensing and signal processing to SensPro, a programmable digital signal processor. NeuPro-Nano handles always-on inference, or running a trained AI model. This division lets each workload run on an engine suited to it, alongside Mobilint’s own processor design.

Mobilint positions REGULUS as a standalone processor that can run Linux and applications without a host computer. Its intended uses include drones, machine vision, smart cameras, robotics and connected devices. That positioning extends beyond adding AI calculations to another machine: the chip also includes functions for handling images, video and device connections.

Those functions include a built-in image processor with high-dynamic-range imaging and noise reduction, plus hardware video encoding at 4K resolution and 60 frames per second. Direct connections support cameras, displays, Ethernet, USB and sensors. These are parts of the chip’s device-processing design, alongside the neural and signal-processing engines.

10 TOPSAI performance

The announcement rates REGULUS at 10 trillion operations per second, with power consumption below 3 watts.

Up to 8GBMemory support

The chip supports up to 8GB of memory, according to the company announcement.

A shipping accelerator and three ways to use the chip

The commercial example is Mobilint’s MLD-R1 AI USB accelerator. First previewed at COMPUTEX 2026, it is now shipping, the announcement says. Unlike REGULUS’s standalone role, this product is designed to add on-device AI to existing systems. Mobilint lists computer vision, video analytics and generative AI among its supported workloads.

Mobilint offers three REGULUS-based formats for different development and deployment needs:

• MLM-1: a system-on-module for adding AI to custom carrier boards.
• REGULUS AI EVK: an evaluation kit for benchmarking and getting software running.
• MLD-R1: a USB-C dongle for adding AI to existing PCs.

Software support provides the route from an existing model to the chip. Mobilint says its qb software development kit compiles PyTorch, TensorFlow and ONNX models for REGULUS. The company also says more than 490 model architectures have been validated on its hardware, spanning convolutional, recurrent and transformer neural networks.

A Yocto Linux development kit supports device software with GStreamer, OpenCV, and camera, video and display drivers. Developers can use the evaluation kit to test workloads and bring up software. Mobilint’s compute rating and model-validation count are company specifications, not performance results for every intended application.