Home > Business > Tech

KAIST-created SoulMate to live on your device as personal AI

Published: 17 Mar. 2026, 15:21 Updated: 17 Mar. 2026, 16:37

Professor Yoo Hoi-Jun’s research team at the KAIST Graduate School of AI Semiconductor presents SoulMate at the International Solid-State Circuits Conference, held in San Francisco, in February. [KAIST]

Korean researchers have developed an artificial intelligence chip capable of on-device model training, paving the way for users to have an AI of their own.

A team led by KAIST professor Yoo Hoi-jun announced Tuesday that it had developed SoulMate, a large language model chip that evolves on its own around each user’s traits and preferences. The researchers described SoulMate as “the world’s first on-device AI capable of both inference and training on the device itself without going through a cloud server.”

Existing AI services have faced problems such as personal data leaks and network delays in the process of sending data to cloud servers. SoulMate’s defining feature is that it can create a deeply personalized AI agent without transmitting information outside the device. In other words, it is designed to help each user build an AI of their own.

The researchers implemented retrieval-augmented generation inside the chip so it can remember a user’s conversations and produce tailored answers. They also built in low-rank adaptation, allowing the system to learn immediately from user feedback.

Where conventional AI has largely stopped at matching training data with user information, SoulMate aims to go a step further by enabling extra training that produces an AI shaped specifically for one person.

“AI such as ChatGPT responds based on the way it was trained in data centers,” Yoo said. “What sets SoulMate apart is that it lets users create an AI of their own, trained only on their own data.”

On-device AI is drawing attention as a promising new technology that could reduce dependence on the cloud. In general, however, large language models require more than 10 billion parameters and more than 8 gigabytes of memory, while devices face hard limits in the amount of power and computing capacity they can support.

A robot dog is on display at CES 2026 Unveiled, held at the Mandalay Bay Convention Center in Nevada, on Jan. 4. [YONHAP]

To improve computational efficiency, the team applied a mixed-rank architecture that assigns different weights according to the importance of tokens, or words. Important content is processed more heavily, while less important content is processed less.

It also reduced energy use during learning from user feedback by identifying similarities between sentences and skipping redundant computations. The researchers reported that SoulMate improved response latency by as much as 82.5 percent and energy efficiency by as much as 76.2 percent over existing technology.

As a result, SoulMate can perform learning and inference at the same time with ultra-low power consumption of 9.8 milliwatts, about one-500th of the power consumed by a smartphone processor. Its response time was 0.2 seconds, or 216.4 milliseconds.

SoulMate could be used in any field that requires personalized AI, including smartphones and wearable devices. If installed on a smartphone, for example, it could create an AI assistant of your own that learns even sensitive information such as a user’s emotions, habits and biometric data.

Its potential combination with physical AI operating in real-world environments is also drawing attention. Interest in the technology’s possible applications is especially strong overseas, particularly in places where privacy is highly valued.

A screen reads "AI in the physical world" as attendees gather during Rivian's first Autonomy and AI Day, showcasing developments in self-driving technology, in Palo Alto, California, Dec. 11, 2025. [REUTERS/YONHAP]

For example, SoulMate could be used to create companion robots capable of nuanced emotional exchanges with users.

“Just as affection grows over time when you raise a dog, emotional interaction with a robot also becomes possible,” Yoo said. “It can even recognize and respond to a user’s speaking style and habits.”

SoulMate was first unveiled at the International Solid-State Circuits Conference (Isscc) held in San Francisco in February. It was also selected as a highlight paper at Isscc, which is often referred to as the Olympics of semiconductor design.

At the time, Neuralink, the neurotechnology company founded by Elon Musk, invited the researchers to its Silicon Valley headquarters for a meeting. The team plans to commercialize SoulMate by 2027 through a startup founded by Yoo.

This article was originally written in Korean and translated by a bilingual reporter with the help of generative AI tools. It was then edited by a native English-speaking editor. All AI-assisted translations are reviewed and refined by our newsroom.
BY CHANG YOON-SEO [[email protected]]