Mobilint ARIES and REGULUS Edge AI NPUs

Mobilint is a South Korean fabless AI chip company established in 2019 and based in Gangnam-gu, Seoul. It designs neural processing units so that AI models, including language models, can run on-premise and on-device instead of in GPU clusters or the cloud, targeting robotics, smart factories, smart cities and server infrastructure.

The ARIES accelerator delivers 80 TOPS with 16 GB LPDDR4X (32 GB optional) over PCIe Gen4 x8 at a 25 W TDP, and is sold as MLA100 and MLA400 PCIe cards, an MXM module and MLX-A1 edge AI boxes. The REGULUS system-on-chip targets compact devices through an MLM-1 system-on-module, an M.2 card, a USB stick and a mini AI box. The qb SDK supports TensorFlow, PyTorch and ONNX, runs up to 32 models concurrently and has been tested on hundreds of CNN, RNN and transformer models, with benchmarks such as Llama-3.2-3B and YOLO models published on the site.

• ARIES NPU: 80 TOPS with 16 GB LPDDR4X and PCIe Gen4 x8 at a 25 W TDP.
• Accelerator cards: MLA100 and MLA400 PCIe cards plus an MXM module for servers and industrial PCs.
• Edge AI boxes: MLX-A1 and MLX-A1D with MIPI, USB and Ethernet connectivity.
• REGULUS SoC: low-power chip offered as SoM, M.2 card, USB stick and mini AI box.
• qb SDK: supports TensorFlow, PyTorch and ONNX and runs up to 32 models concurrently.
• Model Zoo: ready models on GitHub and Hugging Face, including LLM and YOLO benchmarks.

A Vietnamese expressway or city traffic operator can add ARIES cards to video servers to analyse thousands of CCTV streams for incidents. A factory can deploy MLX-A1 boxes at production lines for defect detection. An enterprise or public agency can run a local language model on an ARIES server to keep documents inside its network.

One ARIES card can run many camera streams and models at once, for example ResNet-50 at 3,082 frames per second as published by the company, reducing the number of GPU servers needed for video analytics. Keeping AI on-premise protects sensitive video and documents. The SDK claims 99% of original model accuracy is retained after conversion. A range of form factors lets the same software scale from a USB stick to multi-card servers.

Public agencies and enterprises are moving video analytics and language models on-premise for privacy and cost reasons, and Korea is promoting domestic NPUs through public procurement. Southeast Asian cities and highway operators expanding CCTV networks, and factories adding vision inspection, need dense low-power inference that can sit in existing server rooms and roadside cabinets.

Established in 2019 and headquartered in Seoul, South Korea. Raised seed (2019), Series A (2021) and Series B (2024) rounds. Launched ARIES in 2022 and REGULUS in 2024. Holds ISO 9001 certification and is a founding member of MLCommons. Won CES Innovation Awards for REGULUS (2025) and the MLX-A1 Edge AI Box (2026), and Korean Minister of Science and ICT awards.