Ceva's NeuPro NPU reaches production in Mobilint's REGULUS chip: 10 TOPS under 3W for physical AI

Ceva announced on November 2, 2026 that Mobilint, a fabless AI semiconductor company founded in 2019 and headquartered in Seoul, has deployed Ceva’s SensPro AI DSP and NeuPro-Nano NPU in its REGULUS edge AI system-on-chip. Ceva describes the deployment as the first production deployment of its NeuPro NPU family — the moment its edge-AI intellectual property moves from a product portfolio into commercial silicon.

REGULUS is built for what Ceva calls physical AI: systems that must sense, interpret and respond to their surroundings in real time, on the device, rather than sending everything to the cloud. The chip delivers 10 TOPS of AI performance at under 3W and supports up to 8GB of memory, aimed at power- and space-constrained applications including robotics, drones, machine vision, smart cameras and internet-of-things devices.

Two engines for sensing and inference#

REGULUS uses a multi-engine architecture that assigns each workload to the processing engine best suited to it. SensPro provides programmable processing for sensing and signal-processing workloads, while NeuPro-Nano provides efficient neural processing for always-on inference. Together, the two blocks complement Mobilint’s proprietary AI processing architecture, adding flexible sensing and inference for the varied workloads physical AI devices face, according to the announcement.

Mobilint positions REGULUS as a complete standalone processor rather than an accelerator: it runs Linux and applications without a host computer, Unite.AI reports, citing Mobilint’s product information. Beyond inference, the chip includes a built-in image processor with HDR and noise reduction, hardware video encoding at 4K and 60 frames per second, and direct connections for cameras, displays, Ethernet, USB and sensors. Mobilint’s product page describes dense INT8 compute on a 17mm chip.

Already shipping inside a USB accelerator#

REGULUS is already enabling a commercial product. Mobilint’s MLD-R1 AI USB accelerator, first previewed at COMPUTEX 2026, is now shipping, according to the announcement. Designed to add on-device AI to existing systems, it supports computer vision, video analytics and generative AI workloads for smart factories, robotics, machine vision, security, medical devices and digital signage, with support for PyTorch, ONNX, Hugging Face, TensorFlow, TFLite and Keras.

Mobilint offers REGULUS in three forms, Unite.AI reports: the MLM-1 system-on-module for custom carrier boards, the REGULUS AI EVK evaluation kit for benchmarking and software bring-up, and the MLD-R1 USB-C dongle for existing PCs. On software, Mobilint’s qb SDK compiles PyTorch, TensorFlow and ONNX models for the chip, and the company says more than 490 model architectures across CNN, RNN and transformer families have been validated on its hardware.

“REGULUS brings together the specialized processing capabilities needed to run sophisticated AI workloads directly on the device within very tight power constraints,” said Dongjoo Shin, CEO of Mobilint. Yaron Galitzky, executive vice president and general manager of Ceva’s AI division, called the deployment “an important milestone for Ceva as the first production deployment of our NeuPro NPU family,” and evidence of progress in bringing its AI technology into commercial silicon.

What this settles — and what it does not#

For Ceva, the significance is commercial proof: its IP is, by its own account, in more than 22 billion shipped devices for over 400 customers, and REGULUS is the first production outing for the NeuPro NPU family specifically. For Mobilint, a Seoul fabless company founded in 2019, it puts a named partner’s sensing and inference IP inside a chip that is already shipping in at least one product.

What the announcement does not settle is how REGULUS compares with rival edge chips in real workloads. It gives no price and no manufacturing process node, and its 10 TOPS figure is a specification, not a benchmark result. The test that matters now is whether further NeuPro deployments follow — and how the chip’s power and performance hold up once more devices built on it reach customers. Product information for REGULUS and MLD-R1 is available at Mobilint’s REGULUS page; Ceva’s edge AI portfolio is described on Ceva’s site.